Classification and analysis method of super-low frequency electromagnetic disturbance at middle and low latitudes based on satellite observation

Abstract

The application discloses a method for classifying and analyzing super low frequency electromagnetic disturbance at middle and low latitudes based on satellite observation, which comprises the following steps: obtaining electric field component disturbance values and electron density relative change values through feature processing, constructing a classification feature set, inputting in-situ observation data into a disturbance existence determination model to obtain significant disturbance data, calculating cross-covariance and multivariate mutual information between the significant disturbance data, performing multistage disturbance classification to obtain disturbance results and calculate disturbance indexes, constructing a disturbance factor gradient boosting regression model, and inputting in-situ observation data to be processed into the disturbance factor gradient boosting regression model to obtain disturbance prediction indexes and disturbance influencing factors. The method can efficiently and accurately extract and classify super low frequency electromagnetic disturbance at middle and low latitudes, and can also perform disturbance prediction and quantitative attribution of influencing factors, which has important significance for deepening the research on the dynamic mechanism of the magnetosphere-ionosphere coupling process and improving the space environment monitoring capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of magnetosphere-ionosphere coupling, and particularly relates to a method for classifying and analyzing ELF electromagnetic disturbances at middle and low latitudes based on satellite observations. BACKGROUND

[0002] Magnetosphere-ionosphere coupling is an important part of space weather research. As a key energy and momentum carrier, accurate ELF wave classification and analysis are of great significance to understanding the dynamics of the coupling process. However, the acquisition of waves in this frequency band depends on ground-based magnetometers and observations from a few high-orbit satellites, which have low spatial coverage and difficulty in obtaining in-situ plasma parameters. The characteristics of low-Earth orbit satellites allow for global in-situ detection and provide more comprehensive and irreplaceable data support.

[0003] The existing ELF electromagnetic disturbance classification and analysis method has the following disadvantages: first, most studies focus on the dayside or high-latitude region, and lack a systematic statistical analysis framework for ELF fluctuations at night in the middle and low latitudes; second, traditional methods rely on manual setting of fixed thresholds for disturbance identification and classification, which is highly subjective and inefficient, ignores the influence of electric field disturbances and plasma parameters, and has a high misjudgment rate; in addition, existing technologies fail to effectively integrate multi-source environmental parameters to build a prediction model, which cannot quantitatively assess the influence of different driving factors on disturbance activity, limiting its application potential in actual space weather forecasting. Therefore, the present application proposes a method for classifying and analyzing ELF electromagnetic disturbances at middle and low latitudes based on satellite observations, which replaces manual threshold determination with a machine learning model, uses mutual covariance and multivariate mutual information for multi-level classification, and builds a gradient boosting regression model with factor analysis capability, achieving accurate disturbance identification and classification, predicting disturbance activity, and quantitatively attributing each influencing factor, providing a powerful analysis tool for top ionospheric dynamics research and having important application value in the field of precise space weather modeling and forecasting. SUMMARY

[0004] The present application aims to provide a method for classifying and analyzing ELF electromagnetic disturbances at middle and low latitudes based on satellite observations.

[0005] To achieve the above-mentioned purpose, the present application is implemented according to the following technical solutions:

[0006] The present application comprises the following steps:

[0007] Obtain in-situ observation data of middle and low latitude Earth orbit satellites and perform preprocessing and feature processing to obtain electric field component disturbance values and electron density relative changes; the in-situ observation data includes electric field data, electron density data, and auxiliary data;

[0008] According to the feature processing result, a classification feature set is constructed, and the in-situ observation data is input into a disturbance existence determination model to obtain significant disturbance data;

[0009] The cross-covariance and multivariate mutual information between the significant disturbance data are calculated, and multi-level disturbance classification is performed on the significant disturbance data to obtain a disturbance result; the disturbance result includes synchronous disturbance and non-synchronous disturbance;

[0010] According to the disturbance result and the corresponding in-situ observation data, a disturbance index is calculated, and a disturbance factor gradient boosting regression model is constructed; the in-situ observation data to be processed is input into the disturbance factor gradient boosting regression model to obtain a disturbance prediction index and a disturbance influencing factor.

[0011] Further, the method for obtaining the electric field component disturbance value and the relative change amount of electron density includes:

[0012] In-situ observation data of a medium-low latitude Earth orbit satellite are acquired; the in-situ observation data include electric field data, electron density data, and auxiliary data;

[0013] The electric field data are converted from a satellite coordinate system to a geomagnetic coordinate system to obtain electric field components; the electric field components include a radial component , an azimuthal component , and a parallel component ;

[0014] The length of in-situ data collected by a medium-low latitude Earth orbit satellite in half an orbit is taken as a sliding window length, an adaptive detrending method based on wavelet transform is used to perform disturbance extraction on the electric field data and the electron density data to obtain electric field component disturbance values and electron density disturbance values ; the electric field component disturbance values include radial component disturbance , azimuthal component disturbance , and parallel component disturbance ;

[0015] The ratio of the electron density disturbance value to the original electron density is taken as the relative change amount of electron density .

[0016] Further, the method for constructing a classification feature set includes:

[0017] An ultra-low frequency oscillation region is determined according to the electric field component disturbance values and the auxiliary data; the ultra-low frequency range is 17-100 mHz;

[0018] The power spectral density of each electric field component disturbance value is calculated, and the power spectral density of each electric field component disturbance value in the ultra-low frequency oscillation region is accumulated to obtain the total power of the ultra-low frequency wave of the corresponding component;

[0019] determining the noise reference according to the total power of the parallel component ultra-low frequency wave, and calculating the sum of the total power of the radial component ultra-low frequency wave and the total power of the azimuth component ultra-low frequency wave as the total power of the direction low frequency wave;

[0020] extracting the auxiliary data to obtain environmental features; the environmental features include season encoding, hemispheric position, geomagnetic activity index, and solar activity level;

[0021] The classification feature set is composed of the historical noise reference, the total power of the direction low frequency wave, the environmental features, and the corresponding significant disturbance label.

[0022] Further, the method for obtaining significant disturbance data comprises:

[0023] The disturbance existence determination model is trained using the classification feature set; the disturbance existence determination model comprises an input layer, a preliminary screening layer, a model reasoning layer, and a decision output layer;

[0024] The input layer performs feature splicing and standardization processing on the historical noise reference, the total power of the direction low frequency wave, and the environmental features to obtain standard classification features;

[0025] The preliminary screening layer performs conditional preliminary screening according to the input standard classification features, and inputs the standard classification features that meet the conditional preliminary screening into the model reasoning layer; the conditional preliminary screening specifically refers to screening the standard classification features whose total power of the direction low frequency wave is greater than the historical noise reference and whose environmental features meet the corresponding environmental limit;

[0026] The model reasoning layer learns the complex nonlinear combination relationship of the total power of the direction low frequency wave, the noise reference, the environmental features, and the significant disturbance label using gradient boosting decision tree, and calculates the probability score of the significant disturbance;

[0027] The decision output layer performs Boolean logic decision on the probability score of the significant disturbance, and outputs the significant disturbance determination result of the classification features;

[0028] The in-situ observation data is input into the disturbance existence determination model to obtain the significant disturbance determination result, and the corresponding electric field component disturbance value and electron density relative change amount are taken as the significant disturbance data.

[0029] Further, the method for performing multi-level disturbance classification to obtain the disturbance result comprises:

[0030] The cross-covariance of the electric field component disturbance value and the electron density relative change amount in the significant disturbance data in a specific time delay window is calculated, and the expression is:

[0031]

[0032] wherein is the cross-covariance value at the time delay , is a time delay value range, is an electric field component perturbation value, , is an electric field component set, including a radial component, an azimuthal component and a parallel component, is an electron density relative variation amount, is an electric field component perturbation time series mean value, is an electron density relative perturbation time series mean value;

[0033] Calculate the time delay of the mutual information of the electric field component perturbation value and the electron density relative variation amount, expressed as:

[0034]

[0035] wherein is the mutual information at the time delay , is a time delay standard deviation, is a time window length for mutual information calculation, containing discrete values of multiple groups of electric field component perturbation values and electron density relative variation amounts, is a joint probability distribution of variables and , is a marginal probability distribution of variable , is a marginal probability distribution of variable ;

[0036] Compare the cross-covariance and mutual information at the time delay with the corresponding threshold values, and when the cross-covariance is greater than the cross-covariance threshold value and the mutual information is greater than the mutual information difference threshold value, define the perturbation at the time delay as a synchronous perturbation, and vice versa as an asynchronous perturbation.

[0037] Further, the method for calculating the perturbation index comprises:

[0038] Setting a time window, respectively calculating the proportion of the number of synchronous perturbation half-orbits and the number of asynchronous perturbation half-orbits in the total number of half-orbits in the in-situ observation data in the time window to obtain the perturbation occurrence rate; the perturbation occurrence rate includes a synchronous perturbation occurrence rate and an asynchronous perturbation occurrence rate;

[0039] Respectively calculating the half-orbit average direction ultra-low frequency wave total power of different perturbation types in the time window in-situ observation data; the half-orbit average direction ultra-low frequency wave total power takes the mean value of the half-orbit radial component ultra-low frequency wave total power, the half-orbit azimuthal component ultra-low frequency wave total power and the half-orbit parallel component ultra-low frequency wave total power;

[0040] The disturbance index is composed of the disturbance occurrence rate and the total power of the semi-orbital average direction ultra-low frequency wave.

[0041] Further, the method for obtaining the disturbance prediction index and the disturbance influencing factor comprises:

[0042] The historical disturbance index, the environmental characteristics and the electron density data are combined to form an analysis index set, the analysis index set is randomly divided into a training set and a test set according to a ratio of 6:4, the disturbance factor gradient boosting regression model is trained by using the training set, and the performance of the disturbance factor gradient boosting regression model is evaluated by using the test set.

[0043] The disturbance factor gradient boosting regression model comprises an input layer, a prediction calculation layer, an attribution layer and an output layer.

[0044] The prediction calculation layer is used to fit the complex linear relationship between the disturbance index, the environmental characteristics and the electron density data, and to predict the disturbance index according to the in-situ observation data.

[0045] The attribution layer is used to extract the weighted average value of the total information gain brought by each variable in all decision trees as a disturbance contribution degree, and to determine the disturbance factor according to the disturbance contribution degree of each variable.

[0046] The output layer is connected to the prediction calculation layer and the attribution layer, and outputs the prediction index and the disturbance influencing factor.

[0047] The disturbance factor gradient boosting regression model adopts a multi-objective loss function to improve the prediction accuracy and the stability of the model; the multi-objective loss function comprises a prediction error penalty, a prediction stability constraint and a feature sparsity guide, and the expression is as follows:

[0048]

[0049] wherein is the total loss of the model, is the number of training samples, is the real disturbance index of the i th sample, is the prediction value of the i th sample, is the weight coefficient of the stability penalty term, is the mean value of the prediction values of all samples, is the weight coefficient of the L1 regularization term, is the number of variables, is the weighted mean value of the decision trees related to the i th variable;

[0050] The in-situ observation data to be processed are input into the disturbance factor gradient boosting regression model to obtain the disturbance prediction index and the disturbance influencing factor.​​​

[0051] The beneficial effects of the present application are:

[0052] The present application is based on satellite observation of low-latitude ultra-low frequency electromagnetic disturbance classification and analysis method, compared with the prior art, the present application has the following technical effects:

[0053] The present application can improve the data preprocessing ability and enhance the model adaptability in the classification and analysis of low-latitude ultra-low frequency electromagnetic disturbance, can improve the automation and objectivity of disturbance identification and classification, can improve the precision and physical reliability of disturbance type distinction, can quantitatively analyze the influencing factors of disturbance activity, optimize the classification and analysis method of low-latitude ultra-low frequency electromagnetic disturbance, and constitutes a complete analysis chain from identification, classification to attribution, which can further reveal the dynamics mechanism of magnetosphere-ionosphere coupling process, and provides strong tool support for the mechanism research and application prediction of space weather. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The present application is based on satellite observation of low-latitude ultra-low frequency electromagnetic disturbance classification and analysis method, compared with the prior art, the present application has the following technical effects:

[0055] Figure 2 The feature processing result data relationship diagram of the embodiment of the present application;

[0056] Figure 3 The time ultra-low frequency wave total power relationship diagram of the embodiment of the present application;

[0057] Figure 4 The field component and time offset relative change amount of electron density mutual covariance relationship diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0058] The present application will be further described below through specific embodiments, the illustrative embodiments of the present application and the description are used to explain the present application, but not as a limitation of the present application.

[0059] The present application is based on satellite observation of low-latitude ultra-low frequency electromagnetic disturbance classification and analysis method, compared with the prior art, the present application has the following technical effects:

[0060] As shown in Figure 1 In the present embodiment, the following steps are included:

[0061] Obtain low-latitude earth orbit satellite in-situ observation data and perform preprocessing and feature processing to obtain electric field component disturbance value and electron density relative change amount; the in-situ observation data includes electric field data, electron density data and auxiliary data;

[0062] According to the feature processing result, a classification feature set is constructed, and the in-situ observation data is input into a disturbance existence determination model to obtain significant disturbance data;

[0063] The cross-covariance and multivariate mutual information between the significant disturbance data are calculated, and multi-level disturbance classification is performed on the significant disturbance data to obtain a disturbance result; the disturbance result includes synchronous disturbance and non-synchronous disturbance;

[0064] According to the disturbance result and the corresponding in-situ observation data, a disturbance index is calculated, and a disturbance factor gradient boosting regression model is constructed; the in-situ observation data to be processed is input into the disturbance factor gradient boosting regression model to obtain a disturbance prediction index and a disturbance influencing factor.

[0065] In this embodiment, the method for obtaining the electric field component disturbance value and the relative variation of electron density includes:

[0066] In-situ observation data of a medium-low latitude Earth orbit satellite are acquired; the in-situ observation data include electric field data, electron density data, and auxiliary data;

[0067] The electric field data are converted from a satellite coordinate system to a geomagnetic coordinate system to obtain electric field components; the electric field components include a radial component , an azimuthal component , and a parallel component ;

[0068] The length of the in-situ data collected by the medium-low latitude Earth orbit satellite in half an orbit is taken as the length of a sliding window, and an adaptive detrending method based on wavelet transform is used to extract disturbances from the electric field data and the electron density data to obtain electric field component disturbance values and electron density disturbance values ; the electric field component disturbance values include radial component disturbance , azimuthal component disturbance , and parallel component disturbance ;

[0069] The ratio of the electron density disturbance value to the original electron density is taken as the relative variation of electron density .

[0070] In this embodiment, the method for constructing a classification feature set includes:

[0071] According to the electric field component disturbance values and the auxiliary data, an ultra-low frequency oscillation region is determined; the ultra-low frequency range is 17-100 mHz;

[0072] The power spectral density of each electric field component disturbance value is calculated, and the power spectral densities of each electric field component disturbance value in the ultra-low frequency oscillation region are accumulated to obtain the total power of the ultra-low frequency wave of the corresponding component.

[0073] The sum of the total power of the radial component ultra-low frequency wave and the total power of the azimuthal component ultra-low frequency wave is calculated as the total power of the direction low frequency wave according to the noise reference determined according to the total power of the parallel component ultra-low frequency wave;

[0074] The auxiliary data is subjected to feature extraction to obtain environmental features; the environmental features include season encoding, hemispheric position, geomagnetic activity index and solar activity level;

[0075] The classification feature set is composed of the historical noise reference, the total power of the direction low frequency wave, the environmental features and the corresponding significant disturbance label;

[0076] In the actual evaluation, the electric field data and the electron density data of all DC / ULF frequency bands observed by the XX satellite at a height of 660 km are taken as examples to perform the classification and analysis of the ultra-low frequency electromagnetic disturbance at the middle and low latitudes; the observation time is from January 2006 to November 2010, the XX satellite covers two local times, which are 10:30 LT (day side) and 22:30 LT (night side) respectively, the observation data is organized in the day side or the night side half orbit, the observation range is limited between the invariant latitude-65° and 65°, the satellite coordinate system DC / ULF frequency band electric field waveform data provided by the XX satellite data center has a sampling rate of 39.0625 Hz and a resolution of 40 μV / m, and the time resolution of the electron density observed by the Langmuir probe is 1 s;

[0077] The three components of the electric field are converted from the satellite coordinate system to the local geomagnetic coordinate system to obtain the radial component , the azimuthal component and the parallel component of the electric field, parallel to the earth's magnetic field, pointing east perpendicular to the magnetic meridian plane, pointing to the higher magnetic shell in the magnetic meridian plane;

[0078] The disturbance extraction is performed on the electric field data and the electron density data to obtain the electric field component disturbance value and the electron density disturbance value , and the ratio of the electron density disturbance value to the original electron density is taken as the electron density relative change .

[0079] In the embodiment, the method for constructing the classification feature set comprises:

[0080] The ultra-low frequency oscillation region is determined according to the electric field component disturbance value and the auxiliary data; the ultra-low frequency range is 17-100 mHz;

[0081] The power spectrum density of each electric field component perturbation value is calculated, and the power spectrum density of each electric field component perturbation value in the ultra-low frequency range in the ultra-low frequency oscillation region is accumulated to obtain the total power of the ultra-low frequency wave corresponding to the component;

[0082] The noise reference is determined according to the parallel component total power of the ultra-low frequency wave, and the sum of the radial component total power of the ultra-low frequency wave and the azimuth component total power of the ultra-low frequency wave is calculated as the direction low frequency wave total power;

[0083] The auxiliary data is feature extracted to obtain environmental features; the environmental features include season encoding, hemispheric position, geomagnetic activity index and solar activity level;

[0084] The classification feature set is composed of historical noise reference, direction low frequency wave total power, environmental features and corresponding significant perturbation label;

[0085] In actual evaluation, Figure 2 is a relationship diagram of electric field component perturbation value, electron density relative change and auxiliary data (time position UT, latitude L) drawn according to historical feature processing results, according to Figure 2 Visual inspection of the data shows that the ultra-low frequency oscillation basically occurs on the night side, while the electric field observation data on the day side is very quiet, and two types of ultra-low frequency oscillation occur (a Figure 2 a is the ultra-low frequency oscillation caused by synchronous disturbance, Figure 2 b is the ultra-low frequency oscillation caused by non-synchronous disturbance), in addition, the two types of ultra-low frequency oscillation often occur at middle and low latitudes (L<2), so the middle and low latitudes (L<2) on the night side are determined as the ultra-low frequency oscillation region;

[0086] The power spectrum density of each electric field component perturbation value in the ultra-low frequency range in the ultra-low frequency oscillation region is calculated (c Figure 2 c is the power spectrum density of synchronous disturbance, Figure 2 d is the power spectrum density of non-synchronous disturbance), the specific calculation method is as follows: for each half track, the electric field component perturbation value is divided into continuous and overlapping time periods, each time period lasts about 2 minutes (an electric field sampling sequence contains 256 data points, there are 18 sampling sequences in a 2-minute interval, the actual sampling time of each time period is 118s, and the corresponding frequency resolution is 8.5mHz), and the overlap is 30%, the data of each time period is multiplied by a Hamming window to reduce edge effects, and then the power spectrum density of the electric field component perturbation value of each time period is obtained;

[0087] For each half track, the power spectrum density of 17-100 mHz in the L<2 region is added to obtain the total power of the ultra-low frequency wave of each component, and a time total power of the ultra-low frequency wave relationship diagram is drawn, such as Figure 3 ( Figure 3 a is the 21893 half track data on the night side, Figure 3 b is the 22205 half track data on the day side) according toFigure 3 The ultra-low frequency oscillation is mainly found in the nightside and component, while the total power of the ultra-low frequency wave in the dayside and component is at least one order of magnitude smaller than that in the nightside, and the total power of the ultra-low frequency wave in the nightside and dayside component is basically less than 200 (mV / m) 2 , twice the value is taken as the noise reference (when the nightside is triggered by a significant ultra-low frequency oscillation, the total power of the direction low frequency wave will be significantly greater than the noise reference), and the environmental limit condition (nightside, northern hemisphere summer / southern hemisphere winter, semi-annual cycle) is determined;

[0088] In the environmental characteristics, the seasonal coding is represented by the sine / cosine component, the hemispheric position is a binary variable, and the geomagnetic activity index includes the average value of the nightside substorm activity index AE, the average value of the global geomagnetic activity index Kp, the average value of the ring current intensity index Dst, and the solar activity level in the half-track period.

[0089] The classification feature set is composed of historical noise reference, total power of direction low frequency wave, environmental characteristics and corresponding significant disturbance label.

[0090] In this embodiment, the method for obtaining significant disturbance data comprises:

[0091] The classification feature set is trained to train the disturbance existence determination model; the disturbance existence determination model comprises an input layer, a preliminary screening layer, a model reasoning layer and a decision output layer;

[0092] The input layer performs feature splicing and standardization processing on the historical noise reference, the total power of the direction low frequency wave and the environmental characteristics to obtain standard classification features;

[0093] The preliminary screening layer performs conditional preliminary screening according to the input standard classification features, and inputs the standard classification features meeting the conditional preliminary screening into the model reasoning layer; the conditional preliminary screening specifically refers to screening the standard classification features whose total power of the direction low frequency wave is greater than the historical noise reference and whose environmental characteristics meet the corresponding environmental limit;

[0094] The model reasoning layer learns the complex nonlinear combination relationship of the total power of the direction low frequency wave, the noise reference, the environmental characteristics and the significant disturbance label, and calculates the probability score of the significant disturbance;

[0095] The decision output layer performs Boolean logic decision on the probability score of the significant disturbance, and outputs the significant disturbance determination result of the classification feature;

[0096] The in-situ observation data is input into the disturbance existence determination model to obtain the significant disturbance determination result, and the corresponding electric field component disturbance value and electron density relative change are taken as the significant disturbance data;

[0097] In the actual evaluation, a perturbation existence determination model was trained using supervised learning. The random forest classifier in the ensemble learning algorithm was selected for the model inference layer, and the probability score threshold for the decision output layer was set to 85% (a significant perturbation was determined to be significant when the probability score of a significant perturbation was greater than 85%).

[0098] In this embodiment, the method for obtaining perturbation results by performing multi-level perturbation classification includes:

[0099] The cross-covariance of the electric field component perturbation and the relative change in electron density in significantly perturbed data within a specific time delay window is calculated using the following expression:

[0100]

[0101] in For time delay The cross-covariance value at the location, The range of values ​​for the time delay is [not specified]. This represents the disturbance value of the electric field component. , It is a set of electric field components, including radial, azimuth, and parallel components. This represents the relative change in electron density. electric field The mean of the component perturbation time series, The mean of the relative perturbation time series of electron density;

[0102] Calculate time delay The mutual information between the perturbation value of the electric field component and the relative change in electron density is expressed as:

[0103]

[0104] in For time delay Interactive information at the time The standard deviation of the time delay. The time window length for mutual information calculation includes discrete values ​​of multiple sets of electric field component perturbation values ​​and relative changes in electron density. For variables and The joint probability distribution, For variables The marginal probability distribution, For variables The marginal probability distribution;

[0105] Delay time The cross-covariance and mutual information at each point are compared with their corresponding thresholds. A time delay is defined when the cross-covariance is greater than the cross-covariance threshold and the mutual information is greater than the mutual information difference threshold. A disturbance at a certain point is defined as a synchronous disturbance, and vice versa;

[0106] In practical evaluation, the cross-covariance of the electric field component perturbation value and the relative change of electron density in the historical significant perturbation data within the time delay window (-120, 120) (unit: s) is calculated according to the cross-covariance formula, and the relative change of electron density of the electric field component and the time offset is plotted. Covariance diagram, based on Figure 4 It was found that when there is synchronous perturbation in electric field and electron density, the maximum value of cross covariance appears near time delay 0s. The maximum cross covariance of all asynchronous perturbations is basically less than 1000. Therefore, the cross covariance threshold is set to 1000.

[0107] The cross-covariance formula is used to calculate the cross-covariance of the electric field component perturbation value and the relative change of electron density in the significantly perturbed data within the time delay window (-120, 120) (unit: s).

[0108] The mutual information of electric field component perturbation value and relative change of electron density in historical significant perturbation data within the time window length (-900, 900) (unit: s) is calculated according to the mutual information formula, and the mutual information difference threshold is set to 0.80.

[0109] Delay time The cross-covariance and mutual information at each point are compared with the corresponding thresholds. A time delay is defined when the cross-covariance of all three electric field components and the relative changes in electron density are greater than the cross-covariance threshold, and all mutual information values ​​are greater than the mutual information difference threshold. A disturbance at a certain point is defined as a synchronous disturbance, and vice versa;

[0110] The in-situ observation data of XX satellite at an altitude of 660 km (after feature processing) was classified into multiple levels to obtain the disturbance results. The main locations of synchronous disturbances of electric field and electron density were night side and winter in the Northern Hemisphere, and the cross-covariance of the disturbance values ​​of electric field component and the relative change of electron density were both greater than 1000.

[0111] In this embodiment, the method for calculating the disturbance index includes:

[0112] Set a time window, and calculate the proportion of synchronous and asynchronous half-orbits in the in-situ observation data within the time window to obtain the disturbance occurrence rate; the disturbance occurrence rate includes the synchronous disturbance occurrence rate and the asynchronous disturbance occurrence rate;

[0113] Calculate the half-track average directional ultra-low frequency wave total power of different disturbance types in the time window in-situ observation data respectively; the half-track average directional ultra-low frequency wave total power takes the average of the half-track radial component ultra-low frequency wave total power, the half-track azimuth component ultra-low frequency wave total power and the half-track parallel component ultra-low frequency wave total power;

[0114] The disturbance index is composed of the disturbance occurrence rate and the half-track average directional ultra-low frequency wave total power.

[0115] In the embodiment, the method for obtaining the disturbance prediction index and the disturbance influencing factor comprises:

[0116] The historical disturbance index, the environmental characteristics and the electron density data are combined to form an analysis index set, the analysis index set is randomly divided into a training set and a test set according to a ratio of 6:4, the disturbance factor gradient boosting regression model is trained by using the training set, and the performance of the disturbance factor gradient boosting regression model is evaluated by using the test set.

[0117] The disturbance factor gradient boosting regression model comprises an input layer, a prediction calculation layer, an attribution layer and an output layer.

[0118] The prediction calculation layer adopts the gradient boosting regression tree to fit the complex linear relationship between the disturbance index, the environmental characteristics and the electron density data, and predicts the disturbance index according to the in-situ observation data.

[0119] The attribution layer is used to extract the weighted average value of the total information gain brought by each variable in all decision trees as a disturbance contribution degree, and determine the disturbance factor according to the disturbance contribution degree of each variable.

[0120] The output layer is connected to the prediction calculation layer and the attribution layer, and outputs the prediction index and the disturbance influencing factor.

[0121] The disturbance factor gradient boosting regression model adopts a multi-objective loss function to improve the prediction accuracy and the stability of the model; the multi-objective loss function comprises a prediction error penalty, a prediction stability constraint and a feature sparsity guide, and the expression is as follows:

[0122]

[0123] Wherein, is the total loss of the model, is the number of training samples, is the real disturbance index of the i th sample, is the predicted value of the i th sample, is the weighted coefficient of the stability penalty term, is the mean value of the predicted values of all samples, is the weighted coefficient of the L1 regularization term, ​​as the number of independent variables, as the weight mean of the decision tree related to the first independent variable;

[0124] inputting in-situ observation data to be processed into the disturbance factor gradient boosting regression model to obtain a disturbance prediction index and a disturbance influencing factor;

[0125] In actual evaluation, historical disturbance indicators, environmental characteristics and electron density data are combined to form an analysis index set, wherein the electron density data is O+ density proportion, H+ density and total ion density;

[0126] In the prediction calculation layer, the GBRT algorithm realized by the LightGBM library is used to fit the complex linear relationship between the disturbance indicators, environmental characteristics and electron density data;

[0127] In the attribution layer, the disturbance contributions of each independent variable (environmental characteristics and electron density data) are sorted, and the top 3 independent variables with the highest disturbance contribution are selected as the disturbance factors;

[0128] In the multi-objective loss function, the weight coefficient of the stability penalty term is 0.2, and the weight coefficient of the L1 regularization term is 0.05;

[0129] In-situ observation data (after feature processing) of XX satellite at an altitude of 660 km is input into the disturbance factor gradient boosting regression model to obtain a disturbance prediction index and a disturbance influencing factor:

[0130] Above the region of geomagnetic longitude-180° to 0°, the occurrence rate of synchronous disturbance of electric field and electron density is 15%, the mean total power of radial / azimuthal / parallel component ultra-low frequency wave is 380 / 420 / 180 (mV / m)², and the disturbance influencing factors mainly include latitude (the power is stronger in the region of L>1.2 in the southern hemisphere), geomagnetic longitude (the power is weaker outside the region of (-180°, 0°)) and the mean global geomagnetic activity index Kp;

[0131] The occurrence rate of synchronous disturbance of electric field and electron density is 33.3%, the mean total power of radial / azimuthal / parallel component ultra-low frequency wave is 720 / 690 / 160 (mV / m)², and the disturbance influencing factors mainly include latitude (the mean total power of radial / azimuthal component ultra-low frequency wave is the strongest in the region of L>3), the mean night-side substorm activity index AE and total ion density.

[0132] The above only describes the preferred embodiments of the present application and is not intended to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for classifying and analyzing ultra-low frequency electromagnetic disturbances at middle and low latitudes based on satellite observations, characterized in that, The method comprises the following steps: S1, obtaining in-situ observation data of a medium-low latitude Earth orbit satellite and performing preprocessing and feature processing to obtain electric field component disturbance values and electron density relative variation amounts; the in-situ observation data comprises electric field data, electron density data and auxiliary data; S2, constructing a classification feature set according to the feature processing result, inputting the in-situ observation data into a disturbance existence determination model to obtain significant disturbance data; S3, calculating cross-covariance and multivariate mutual information between the significant disturbance data, and performing multi-level disturbance classification on the significant disturbance data to obtain a disturbance result; the disturbance result comprises synchronous disturbance and non-synchronous disturbance; S4, calculating a disturbance index according to the disturbance result and corresponding in-situ observation data, constructing a disturbance factor gradient boosting regression model, and inputting to-be-processed in-situ observation data into the disturbance factor gradient boosting regression model to obtain a disturbance prediction index and a disturbance influencing factor; The method for calculating the cross-covariance and the multivariate mutual information between the significant disturbance data comprises: The cross-covariance of the electric field component disturbance values and the electron density relative variation amounts in the significant disturbance data in a specific time delay window is calculated, and the expression is: wherein is a time delay is a cross-covariance value at the time delay is a range of values for the time delay is a perturbation value of the electric field component is a set of electric field components including a radial component, an azimuthal component, and a parallel component is a relative variation of electron density is an electric field is a mean of the electric field component perturbation time series is a mean of the relative perturbation time series of electron density Computing time delays The mutual information of the electric field component perturbation value and the relative variation of electron density is expressed as: wherein is the time delay is the mutual information at time is the time delay standard deviation is the time window length for mutual information calculation, containing discrete values of the perturbation values of the electric field components and the relative variation of the electron density is the joint probability distribution of variables and is the marginal probability distribution of variable is the marginal probability distribution of variable ​​​ 2. The method according to claim 1, wherein the method is characterized by, The method for obtaining the electric field component disturbance values and the electron density relative variation amounts comprises: Obtaining in-situ observation data of a medium-low latitude Earth orbit satellite; the in-situ observation data comprises electric field data, electron density data and auxiliary data; converting the electric field data from a satellite coordinate system to a geomagnetic coordinate system to obtain electric field components; the electric field components include a radial component , an azimuthal component , and a parallel component ; The in-situ data collected by a middle-low latitude earth orbit satellite in a half orbit is taken as a sliding window length, and an adaptive detrending method based on wavelet transform is used to extract the electric field data and the electron density data to obtain electric field component perturbation values and electron density perturbation values ; the electric field component perturbation values include radial component perturbation , azimuth component perturbation and parallel component perturbation ; Computing electron density perturbation values as a relative change in electron density .​ 3. The method according to claim 1, characterized in that, The method for constructing the classification feature set comprises: Determining an ultra-low frequency oscillation region according to the electric field component disturbance values and the auxiliary data; the ultra-low frequency range is 17-100 mHz; Calculating the power spectral density of each electric field component disturbance value, and accumulating the power spectral density of each electric field component disturbance value in the ultra-low frequency oscillation region to obtain the total power of the ultra-low frequency wave of the corresponding component; Determining a noise reference according to the total power of the parallel component ultra-low frequency wave, and calculating the sum of the total power of the radial component ultra-low frequency wave and the total power of the azimuth component ultra-low frequency wave as the total power of the directional low frequency wave; Extracting features from the auxiliary data to obtain environmental features; the environmental features comprise season encoding, hemispheric position, geomagnetic activity index and solar activity level; The classification feature set is composed of historical noise reference, total power of directional low frequency wave, environmental features and corresponding significant disturbance labels.

4. The method according to claim 1, characterized in that, The method for obtaining the significant disturbance data comprises: Training a disturbance existence determination model using the classification feature set; the disturbance existence determination model comprises an input layer, a preliminary screening layer, a model reasoning layer and a decision output layer; The input layer performs feature splicing and standardization processing on the historical noise reference, the total power of the directional low frequency wave and the environmental features to obtain standard classification features; The preliminary screening layer performs conditional preliminary screening according to the input standard classification features, and inputs the standard classification features meeting the conditional preliminary screening into the model reasoning layer; the conditional preliminary screening specifically refers to screening the standard classification features whose total power of the directional low frequency wave is greater than the historical noise reference and whose environmental features meet the corresponding environmental limit; The model reasoning layer learns the complex nonlinear combination relationship of the total power of the directional low frequency wave, the noise reference, the environmental features and the significant disturbance label, and calculates the probability score of the significant disturbance. The decision output layer performs a Boolean logic decision on the probability score of the significant disturbance, and outputs a significant disturbance determination result of the classification feature. The in-situ observation data is input into the disturbance existence determination model to obtain the significant disturbance determination result, and the corresponding electric field component disturbance value and electron density relative change are taken as the significant disturbance data.

5. The method according to claim 1, wherein the method is characterized by, The method for performing multi-level disturbance classification to obtain a disturbance result comprises: delaying the time The cross-covariance and cross-information at the time delay are compared to corresponding threshold values, and when the cross-covariance is greater than the cross-covariance threshold and the cross-information is greater than the cross-information difference threshold, the disturbance at the time delay 6. The method according to claim 1, wherein the method is characterized by, The method for calculating a disturbance index comprises: A time window is set, and the proportions of synchronous disturbance and non-synchronous disturbance half-orbits in the total number of half-orbits in the time window in-situ observation data are calculated to obtain a disturbance occurrence rate; the disturbance occurrence rate comprises a synchronous disturbance occurrence rate and a non-synchronous disturbance occurrence rate; The half-orbit average directional ultra-low frequency wave total power of different disturbance types in the time window in-situ observation data is calculated; the half-orbit average directional ultra-low frequency wave total power is the average of the half-orbit radial component ultra-low frequency wave total power, the half-orbit azimuth component ultra-low frequency wave total power and the half-orbit parallel component ultra-low frequency wave total power; The disturbance index is composed of the disturbance occurrence rate and the half-orbit average directional ultra-low frequency wave total power.

7. The method according to claim 1, wherein the method is characterized by, The method for obtaining a disturbance prediction index and a disturbance influencing factor comprises: The historical disturbance index, environmental characteristics and electron density data are combined to form an analysis index set, the analysis index set is randomly divided into a training set and a test set according to a ratio of 6:4, a disturbance factor gradient boosting regression model is trained using the training set, and the performance of the disturbance factor gradient boosting regression model is evaluated using the test set; The disturbance factor gradient boosting regression model comprises an input layer, a prediction calculation layer, an attribution layer and an output layer; The prediction calculation layer uses a gradient boosting regression tree to fit the complex linear relationship between the disturbance index, the environmental characteristics and the electron density data, and predicts the disturbance index according to the in-situ observation data; The attribution layer is used to extract the weighted average value of the information gain of each variable in all decision trees as a disturbance contribution degree, and determine a disturbance factor according to the disturbance contribution degree of each variable; The output layer is connected to the prediction calculation layer and the attribution layer, and outputs a prediction index and a disturbance influencing factor; The disturbance factor gradient boosting regression model uses a multi-objective loss function to improve the prediction accuracy and the stability of the model; the multi-objective loss function comprises a prediction error penalty, a prediction stability constraint and a feature sparsity guide, and the expression is as follows: wherein is the total loss of the model, is the number of training samples, is the true perturbation indicator of the th sample, is the predicted value of the th sample, is the weight coefficient of the stability penalty term, is the mean of the predicted values of all samples, is the weight coefficient of the L1 regularization term, is the number of independent variables, is the weight mean of the decision tree related to the th independent variable; The disturbance prediction index and the disturbance influencing factor are obtained by inputting the to-be-processed in-situ observation data into the disturbance factor gradient boosting regression model.

Images