CN109100759B - Ionosphere amplitude flicker detection method based on machine learning - Google Patents

Abstract

The invention discloses an ionospheric amplitude flicker detection method combined with machine learning. In the process, firstly, the received GPS signal is processed and calculated to obtain the maximum value and the average value of the amplitude flicker index S4, which are used as learning samples to mark the corresponding flicker event, and the setting label is 1 to indicate that the flicker event occurs, and-1 to indicate that the flicker event does not occur. And then, inputting the samples into an SVM classifier for learning to obtain an optimal classifier. When a new scintillation event feature vector enters the SVM classifier, it will be automatically classified. The detection method can simultaneously detect a large number of scintillation events, improves the detection efficiency, simplifies the detection process and obtains higher and stable accuracy.

Description

Ionosphere amplitude flicker detection method based on machine learning

Technical Field

The invention relates to the field of wireless communication, in particular to an ionosphere amplitude flicker detection method based on machine learning.

Background

With the increasingly wide application of GNSS and other satellite technologies in modern society, ionospheric scintillation and its influence on radio frequency signals have become important parts to be ignored. Ionospheric flicker is caused by irregularities in the ionospheric plasma and refers to rapid fluctuations in amplitude and phase of the radio frequency signal (e.g., GNSS) propagating in the ionosphere. There are many reasons for this phenomenon, including but not limited to solar activity, magnetic storms, local electric fields, electrical conductivity, wave interactions, etc. More specifically, flicker in high latitudes is often associated with solar activity and periods of magnetic storm, and thus flicker prediction and modeling is very difficult. In equatorial and low latitude areas, equatorial ionization anomalies and ionospheric bubbles formed after sunset are potential causes of scintillation events. The flicker affects all spatial radio signals penetrating the ionosphere and may lead to performance degradation of accuracy and continuity. Strong flicker can seriously affect signal acquisition and tracking of GNSS receivers, resulting in loss of lock and navigation failure. Therefore, the accurate and efficient ionospheric scintillation detection method is sought, which is not only beneficial to designing a receiver with better performance so as to improve the positioning precision and reduce the influence of ionospheric scintillation, but also can provide help for establishing ionospheric and space weather models.

ConventionalIonospheric scintillation detection is a manual identification by researchers observing the detection data, a process that is time consuming and impossible to traverse through all data. Therefore, an automatic flicker detection method is proposed to improve the detection efficiency. The most familiar empirical detection method is to determine the amplitude flicker index S₄If the average value is greater than 0.2, a flicker event occurs. There are also more complex detection algorithms based on NP detection theory, such as wavelet decomposition, that distinguish between scintillation and non-scintillation events by assuming that the wavelet coefficients obey a gaussian distribution; there is another complementary integrated empirical mode decomposition method that detects flicker by an accurate measurement of carrier-to-noise ratio. Although the above algorithm is capable of detecting scintillation events, it is limited by non-optimal empirical thresholds or ideal hypothesis models during the design process.

Disclosure of Invention

The purpose of the invention is as follows: aiming at the problems in the prior art, the invention provides an ionospheric amplitude flicker detection method based on machine learning. Compared with other methods, the method has the advantages that the detection accuracy is effectively improved, the detection steps are simplified, mass data can be processed simultaneously, and the detection efficiency is effectively improved.

The technical scheme is as follows: the ionospheric amplitude flicker detection method based on machine learning comprises the following steps:

(1) for data collected by GPS receivers at different positions, data are obtained by adopting a mobile window with the window length of a second and b seconds of movement each time, and the amplitude flicker index S is calculated for the data obtained by the mobile window each time₄(ii) a Wherein a and b are positive integers, and a>b；

(2) Dividing data acquired by GPS receivers at different positions by taking each t minutes as a data block, taking the maximum value and the average value of amplitude flicker indexes in each data block as the characteristic quantity of the data block, and marking whether the data block generates a flicker event or not by adopting a label, wherein t is a positive integer;

(3) taking the characteristic quantity and the corresponding label of part of the data blocks as training samples, dividing the training samples into samples with flicker events and samples without flicker events according to the label, and taking the characteristic quantity and the corresponding label of the rest data blocks as verification samples;

(4) establishing a linear SVM classifier model;

(5) inputting the two types of training samples into a linear SVM classifier model for cross validation to obtain an optimal hyper-parameter in the SVM classifier model and obtain an optimal SVM classifier;

(6) inputting the characteristic quantity in the check sample into an optimal SVM classifier for classification, comparing the output classification result with a corresponding label, and considering the SVM classifier to be qualified when the accuracy reaches a preset value;

(7) and inputting the unknown characteristic quantity of the flicker event data into a qualified SVM classifier, wherein the output of the SVM classifier is the classification result.

Further, the step (1) specifically comprises:

(1.1) multiplying the digital intermediate frequency signal received by the GPS receiver with the local orthogonal carrier signal to generate I and Q baseband signals, and multiplying the I and Q baseband signals with the instantaneous code to obtain I_pAnd Q_p；

(1.2) according to I_pAnd Q_pThe wideband power and the narrowband power of the signal are calculated using the following equations:

in the formula I_p,i、Q_p,iAre respectively a pair I_p、Q_pThe ith sampling value obtained by sampling once every h millisecond, delta t is a power calculation interval time value, and J represents that I is equal to_pAnd Q_pAll the sampling values are divided into J segments, W, at intervals of Δ t_BP,j、N_BP,jRespectively representing the broadband power and the narrowband power of the jth segmented signal;

(1.3) calculating the normalized signal strength of the signal from the wide-band power and the narrow-band power of the signal using the following equation:

S_I,raw,j＝N_BP,j-W_BP,j

in the formula, S_I,norm,jIndicating the normalized signal strength, S, of the jth segmented signal_I,trend,jIs represented by S_I,raw,jThe detrended signal intensity obtained by fitting a polynomial to the 4 th order of (1);

(1.4) adopting a moving window with the window length of a seconds and moving for b seconds each time to acquire data, and calculating the amplitude flicker index S of the data acquired by moving the window each time₄Wherein the amplitude flicker index value S of the data acquired by the k-th moving window₄Comprises the following steps:

in the formula (I), the compound is shown in the specification,

a normalized total set of signal strengths representing a seconds of data in k moving windows, a 1000/Δ t normalized signal strength data, E [ ·]Representing a mathematical expectation.

Further, the step (2) specifically comprises:

(2.1) carrying out non-overlapping division on data acquired by GPS receivers at different positions by taking the data as a data block every t minutes;

(2.2) obtaining a plurality of amplitude flicker indexes S calculated according to data in each data block₄Extracting the maximum value S therefrom_4,maxAnd the mean value S_4,avgAs the characteristic quantity of the data block, and whether the data block generates a flicker event is marked by a label, which is expressed by the following mathematical form:

characteristic amount:

labeling:

in the formula, l represents a data block sequence number,

representing a two-dimensional vector space.

Further, the linear SVM classifier model established in step (4) is:

constraint conditions are as follows:

wherein w is the parameter matrix to be solved, b is the parameter to be solved,

is a one-dimensional space, ξ_lIs the relaxation variable of the ith training sample, m is the number of training samples, C is a hyperparameter representing the tolerance to training samples that exceed the maximized boundary, x^(l)、y^(l)Respectively representing the characteristic quantity and the label of the ith training sample.

Further, the step (5) specifically comprises:

(5.1) introduction of Lagrangian multiplier α_l,β_lThe linear SVM classifier model is represented as:

respectively make L to w_i,b_i,ξ_iDerivation and 0 obtaining:

(5.2) substituting the result into the model established in the step (4), converting into a dual form according to the strong dual relation, and removing the negative sign to obtain the model:

constraint conditions are as follows:

hiding conditions:

solving the model in MATLAB by using a function quadprog to obtain the optimal value w of w₀(ii) a And according to w₀B is solved to obtain the optimal value b₀＝y^(s)-w₀ ^Tx^(s)Wherein ξ_s＝0，x^(s)Is a support vector of α_lTraining sample feature quantity, y corresponding to item not equal to 0^(s)Is a corresponding label;

(5.3) recording the training sample characteristic quantity as X ═ X⁽¹⁾,x⁽²⁾,...,x^(m)) I.e. a 2 x m matrix; the label is denoted as Y ═ Y⁽¹⁾,y⁽²⁾,...,y^(m)) I.e. a 1 × m row vector, combining the above matrices and vectors into a 3 × m matrix Z ═ X; y) as a sample input matrix;

(5.5) integrally inputting the sample input matrix into an SVM classifier model, setting the value of cross validation fold number and the value of the super parameter C, and in the training process, randomly equally dividing the number of input samples into u parts, wherein each u-1 part is used for learning the model, the remaining 1 part is used for testing the learned model to obtain the testing accuracy, and obtaining the average testing accuracy corresponding to the current super parameter C after carrying out u times of training in sequence;

(5.6) changing the value of the hyper-parameter C, and returning to execute the step (5.5), thereby obtaining the average test accuracy corresponding to different values of the hyper-parameter C;

and (5.7) comparing the average accuracy rates corresponding to all the hyper-parameters, finding out the hyper-parameter C corresponding to the maximum accuracy rate as the optimal hyper-parameter, and training under the parameter setting to obtain the model as the optimal classification model.

Has the advantages that: compared with the prior art, the invention has the following remarkable advantages: the invention provides an ionospheric amplitude flicker detection method based on machine learning, which comprises the steps of firstly extracting and marking characteristics from detected data of whether flicker events occur or not, arranging the characteristics into samples, inputting the samples into an established SVM classifier model for learning and finding out an optimal classifier model, and then testing the learned classifier, so that higher accuracy can be obtained, wherein the classifier model is superior to other methods. And finally, the classifier is applied to new scintillation events, classification detection can be automatically carried out, and compared with the traditional method for judging whether scintillation occurs according to the scintillation index S4, the method has higher accuracy, can detect the scintillation event with the scintillation index lower than 0.2, and has important significance for ionosphere structure model and scintillation mechanism research, and the method can process mass data simultaneously, thereby improving the detection efficiency to a great extent.

Drawings

FIG. 1 is a flow chart of one embodiment of the present invention.

Detailed Description

As shown in fig. 1, the ionospheric amplitude flicker detection method based on machine learning according to the present invention includes:

(1) for data collected by GPS receivers at different positions, a moving window with the window length of 10 seconds and moving for 1 second each time is adopted to obtain the data, and the data obtained by the moving window each time is subjected to amplitude flicker index S calculation₄。

The step (1) specifically comprises the following steps:

(1.1) multiplying the digital intermediate frequency signal received by the GPS receiver with the local orthogonal carrier signal to generate I and Q baseband signals, and multiplying the I and Q baseband signals with the instantaneous code to obtain I_pAnd Q_p；

(1.2) according to I_pAnd Q_pThe wideband power and the narrowband power of the signal are calculated using the following equations:

in the formula I_p,i、Q_p,iAre respectively a pair I_p、Q_pTaking the ith sampling value obtained by sampling every 1ms, taking delta t as a power calculation interval time value, taking 20ms, and J to represent that I is equal to_pAnd Q_pAre divided into J segments, W, at intervals of 20ms_BP,j、N_BP,jRespectively representing the broadband power and the narrowband power of the jth 20ms signal;

(1.3) calculating the normalized signal strength of the signal from the wide-band power and the narrow-band power of the signal using the following equation:

S_I,raw,j＝N_BP,j-W_BP,j

in the formula, S_I,norm,jIndicating the normalized signal strength, S, of the jth segmented signal_I,trend,jIs represented by S_I,raw,jThe detrended signal intensity obtained by fitting a polynomial to the 4 th order of (1);

(1.4) adopting a moving window with the window length of 10 seconds and moving for 1 second each time to obtain data, and calculating the amplitude flicker index S of the data obtained by moving the window for each time₄The window can be moved 10 times in 10S, and the flicker index value S can be calculated 1 time every 1 second₄。

Amplitude flicker index value S of data acquired by k-th moving window₄Comprises the following steps:

in the formula (I), the compound is shown in the specification,

represents a total set of normalized signal strengths of a seconds of data located in k moving windows, 10 × 1000/20 ═ 500 normalized signal strength data, E [ · in total]Representing a mathematical expectation.

(2) Dividing data acquired by GPS receivers at different positions into data blocks in every 3 minutes, taking the maximum value and the average value of amplitude flicker indexes in each data block as characteristic quantities of the data block, and marking whether a flicker event occurs in the data block by adopting a label.

The step (2) specifically comprises the following steps:

(2.1) carrying out non-overlapping division on data acquired by GPS receivers at different positions by taking the data as a data block every 3 minutes;

(2.2) obtaining a plurality of amplitude flicker indexes S calculated according to data in each data block₄Extracting the maximum value S therefrom_4,maxAnd the mean value S_4,avgAs the characteristic quantity of the data block, and whether the data block generates a flicker event is marked by a label, which is expressed by the following mathematical form:

characteristic amount:

labeling:

in the formula, l represents a data block sequence number,

representing a two-dimensional vector space.

(3) And taking the characteristic quantity and the corresponding label of 80% of the data blocks as training samples, dividing the training samples into samples with flicker events and samples without flicker events according to the labels, and taking the characteristic quantity and the corresponding label of the rest data blocks as verification samples.

(4) Establishing a linear SVM classifier model y-w^Tx + b, find the parameter w₀And b₀So that y is equal to w₀ ^Tx+b₀The samples are divided into two types as a plane, the samples closest to two sides of the hyperplane are guaranteed to have the farthest distance, and the sample points with the characteristics are the Support Vectors (SV). The above-described maximum boundary problem is expressed in a series of transformations as the following mathematical form:

constraint conditions are as follows:

wherein w is the parameter matrix to be solved, b is the parameter to be solved,

is a one-dimensional space, ξ_lIs the relaxation variable of the ith training sample, m is the number of training samples, C is a hyperparameter representing the tolerance to training samples that exceed the maximized boundary, x^(l)、y^(l)Respectively representing the characteristic quantity and the label of the ith training sample.

(5) Inputting the two types of training samples into a linear SVM classifier model for cross validation to obtain the optimal hyper-parameter in the SVM classifier model and obtain the optimal SVM classifier.

The step (5) specifically comprises the following steps:

(5.1) introduction of Lagrangian multiplier α_l,β_lThe linear SVM classifier model is represented as:

respectively make L to w_i,b_i,ξ_iDerivation and 0 obtaining:

(5.2) substituting the result into the model established in the step (4), converting into a dual form according to the strong dual relation, and removing the negative sign to obtain the model:

constraint conditions are as follows:

hiding conditions:

the above problem is converted into a quadratic programming problem α_lThe method can be solved by using a function quadprog in MATLAB, and the w meeting the maximized boundary problem is obtained by substituting the function quadprog into the hidden condition₀In addition, the characteristics of the Lagrangian function determine α_lThe ith training sample corresponding to item not equal to 0 is the support vector SV (in x)^(s)Express), corresponding can find b₀＝y^(s)-w₀ ^Tx^(s)Wherein ξ_s＝0。

(5.3) the maximization problem is sample learned by the Classification Learner model in MATLAB to find the best classifier. Recording the training sample characteristic quantity as X ═ X⁽¹⁾,x⁽²⁾,...,x^(m)) I.e. a 2 x m matrix; the label is denoted as Y ═ Y⁽¹⁾,y⁽²⁾,...,y^(m)) I.e. a 1 × m row vector, combining the above matrices and vectors into a 3 × m matrix Z ═ X; y) as a sample input matrix;

(5.5) integrally inputting the matrix into a Classification classifier in MATLAB by using a sample input matrix, selecting a linear SVM classifier model, setting a cross validation fold number and a value of a hyper-parameter C, randomly dividing the number of input samples into u parts in the training process, wherein each u-1 part is used for learning the model, 1 part is left for testing the learned model to obtain the testing accuracy, and obtaining the average testing accuracy corresponding to the current hyper-parameter C after carrying out u times of training in sequence;

(5.6) changing the value of the hyper-parameter C, and returning to execute the step (5.5), thereby obtaining the average test accuracy corresponding to different values of the hyper-parameter C; altering the value of the hyperparameter C may approximate 3-fold longer speed setting parameter C values, such as 0.001,0.003,0.01,0.03, 0.1,0.3,1,3,10, 30;

and (5.7) comparing the average accuracy rates corresponding to all the hyper-parameters, finding out the hyper-parameter C corresponding to the maximum accuracy rate as the optimal hyper-parameter, and training under the parameter setting to obtain the model as the optimal classification model.

(6) And inputting the characteristic quantity in the check sample into an optimal SVM classifier for classification, comparing the output classification result with the corresponding label, and considering the SVM classifier to be qualified when the accuracy reaches a preset value.

(7) And inputting the unknown characteristic quantity of the flicker event data into a qualified SVM classifier, wherein the output of the SVM classifier is the classification result. And the operation result is the label of the new flashing event, if the operation result is 1, the occurrence of the flashing event is judged, and if the operation result is-1, the non-occurrence of the flashing event is judged.

According to the invention, the data samples are selected differently, the accuracy rate can generate smaller deviation, and through test statistics, the accuracy rate can reach 96%, and compared with the traditional method, the detection accuracy is improved.

While the invention has been described in connection with what is presently considered to be the most practical and preferred embodiment, it is to be understood that the invention is not to be limited to the disclosed embodiment, but on the contrary, is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims (5)

1. An ionospheric amplitude flicker detection method based on machine learning, comprising:

(1) for data collected by GPS receivers at different positions, data are obtained by adopting a mobile window with the window length of a second and b seconds of movement each time, and the amplitude flicker index S is calculated for the data obtained by the mobile window each time₄(ii) a Wherein a and b are positive integers, and a>b；

(2) Dividing data acquired by GPS receivers at different positions by taking each t minutes as a data block, taking the maximum value and the average value of amplitude flicker indexes in each data block as the characteristic quantity of the data block, and marking whether the data block generates a flicker event or not by adopting a label, wherein t is a positive integer;

(3) taking the characteristic quantity and the corresponding label of part of the data blocks as training samples, dividing the training samples into samples with flicker events and samples without flicker events according to the label, and taking the characteristic quantity and the corresponding label of the rest data blocks as verification samples;

(4) establishing a linear SVM classifier model;

(5) inputting the two types of training samples into a linear SVM classifier model for cross validation to obtain an optimal hyper-parameter in the SVM classifier model and obtain an optimal SVM classifier;

(6) inputting the characteristic quantity in the check sample into an optimal SVM classifier for classification, comparing the output classification result with a corresponding label, and considering the SVM classifier to be qualified when the accuracy reaches a preset value;

(7) and inputting the unknown characteristic quantity of the flicker event data into a qualified SVM classifier, wherein the output of the SVM classifier is the classification result.

2. The machine learning-based ionospheric amplitude flicker detection method of claim 1, wherein: the step (1) specifically comprises the following steps:

(1.1) multiplying the digital intermediate frequency signal received by the GPS receiver with the local orthogonal carrier signal to generate I and Q baseband signals, and multiplying the I and Q baseband signals with the instantaneous code to obtain I_pAnd Q_p；

(1.2) according to I_pAnd Q_pThe wideband power and the narrowband power of the signal are calculated using the following equations:

in the formula I_p,i、Q_p,iAre respectively a pair I_p、Q_pAt the ith sample value, Δ, sampled every h millisecondst is the power calculation interval time value, J represents I_pAnd Q_pAll the sampling values are divided into J segments, W, at intervals of Δ t_BP,j、N_BP,jRespectively representing the broadband power and the narrowband power of the jth segmented signal;

(1.3) calculating the normalized signal strength of the signal from the wide-band power and the narrow-band power of the signal using the following equation:

S_I,raw,j＝N_BP,j-W_BP,j

in the formula, S_I,norm,jIndicating the normalized signal strength, S, of the jth segmented signal_I,trend,jIs represented by S_I,raw,jThe detrended signal intensity obtained by fitting a polynomial to the 4 th order of (1);

(1.4) adopting a moving window with the window length of a seconds and moving for b seconds each time to acquire data, and calculating the amplitude flicker index S of the data acquired by moving the window each time₄Wherein the amplitude flicker index value S of the data acquired by the k-th moving window₄Comprises the following steps:

in the formula (I), the compound is shown in the specification,

a normalized total set of signal strengths representing a seconds of data in k moving windows, a 1000/Δ t normalized signal strength data, E [ ·]Representing a mathematical expectation.

3. The machine learning-based ionospheric amplitude flicker detection method of claim 1, wherein: the step (2) specifically comprises the following steps:

(2.1) carrying out non-overlapping division on data acquired by GPS receivers at different positions by taking the data as a data block every t minutes;

(2.2) obtaining a plurality of amplitude flicker indexes S calculated according to data in each data block₄Extracting the maximum value S therefrom_4,maxAnd the mean value S_4,avgAs the characteristic quantity of the data block, and whether the data block generates a flicker event is marked by a label, which is expressed by the following mathematical form:

characteristic amount:

labeling:

in the formula, l represents a data block sequence number,

representing a two-dimensional vector space.

4. The machine learning-based ionospheric amplitude flicker detection method of claim 1, wherein: the linear SVM classifier model established in the step (4) is as follows:

constraint conditions are as follows:

wherein w is the parameter matrix to be solved, b is the parameter to be solved,

is a one-dimensional space, ξ_lFor the relaxation variable of the ith training sample, m is the number of training samples, C is the hyperparameter representing the relaxation variable for training samples that exceed the maximization boundaryTolerance, x^(l)、y^(l)Respectively representing the characteristic quantity and the label of the ith training sample.

5. The machine learning-based ionospheric amplitude flicker detection method of claim 4, wherein: the step (5) specifically comprises the following steps:

(5.1) introduction of Lagrangian multiplier α_l,β_lThe linear SVM classifier model is represented as:

respectively make L to w_i,b_i,ξ_iDerivation and 0 obtaining:

(5.2) substituting the result into the model established in the step (4), converting into a dual form according to the strong dual relation, and removing the negative sign to obtain the model:

constraint conditions are as follows:

hiding conditions:

solving the model in MATLAB by using a function quadprog to obtain the optimal value w of w₀(ii) a And according to w₀B is solved to obtain the optimal value b₀＝y^(s)-w₀ ^Tx^(s)Wherein x is^(s)Corresponding ξ_s＝0，x^(s)Is a support vector of α_lTraining sample feature quantity, y corresponding to item not equal to 0^(s)Is a corresponding label;

(5.3) recording the training sample characteristic quantity as X ═ X⁽¹⁾,x⁽²⁾,...,x^(m)) I.e. a 2 x m matrix; the label is denoted as Y ═ Y⁽¹⁾,y⁽²⁾,...,y^(m)) I.e. a 1 × m row vector, combining the above matrices and vectors into a 3 × m matrix Z ═ X; y) as a sample input matrix;

(5.5) integrally inputting the sample input matrix into an SVM classifier model, setting the value of cross validation fold number and the value of the super parameter C, and in the training process, randomly equally dividing the number of input samples into u parts, wherein each u-1 part is used for learning the model, the remaining 1 part is used for testing the learned model to obtain the testing accuracy, and obtaining the average testing accuracy corresponding to the current super parameter C after carrying out u times of training in sequence;

(5.6) changing the value of the hyper-parameter C, and returning to execute the step (5.5), thereby obtaining the average test accuracy corresponding to different values of the hyper-parameter C;

and (5.7) comparing the average accuracy rates corresponding to all the hyper-parameters, finding out the hyper-parameter C corresponding to the maximum accuracy rate as the optimal hyper-parameter, and training under the parameter setting to obtain the model as the optimal classification model.

Images