Strong ionospheric scintillation detection method based on resampling improved gradient boosting tree

Through the gradient enhancement tree method based on resampling, combined with SMOTE-ENN technology and XGBoost algorithm, the problem of classification imbalance in ionosphere scintillation detection is solved, and the detection accuracy and efficiency of weak, medium and strong ionosphere scintillation events are improved. , especially the recognition ability of strong ionospheric scintillation events has been significantly improved.

CN114897010BActive Publication Date: 2024-11-08SOUTHEAST UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202210464770.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-29
Publication Date
2024-11-08
Estimated Expiration
2042-04-29

AI Technical Summary

Technical Problem

In the ionosphere scintillation detection, the sample proportion of strong ionosphere scintillation events is small, resulting in model training deviations, affecting the detection accuracy. The traditional methods lack the flexibility and accuracy of weak, medium and strong ionosphere scintillation events.

Method used

Using a gradient enhancement tree method based on resampling, the training set is resampled through SMOTE-ENN technology, combined with the XGBoost algorithm, multiple decision trees are generated to improve detection accuracy and efficiency, and ensure weak, medium and strong ionosphere flickering events. Automatic detection.

Benefits of technology

It effectively solves the missed detection problem caused by classification imbalance, improves the detection accuracy and efficiency of ionosphere scintillation events, and especially significantly improves the recognition ability of strong ionosphere scintillation events.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114897010B_ABST
    Figure CN114897010B_ABST
Patent Text Reader

Abstract

A strong ionospheric scintillation detection method based on resampling improved gradient boosting tree is proposed. This method first extracts the correlation values ​​of the I and Q branches in the GNSS signal and calculates the mathematical expressions related to them as the feature vectors of the ionospheric scintillation detection model. The scintillation events are divided into three levels of weak, medium and strong by manual classification and marked. The preprocessed data set is divided into a training set and a test set according to a certain ratio. The training set is resampled using the SMOTE‑ENN technology to obtain a sample-balanced data set for modeling and optimization of the XGBoost algorithm, and the test set is used for model performance evaluation. The trained classification model can realize the classification of ionospheric scintillation events under various scintillation intensities. Compared with the traditional method, this method effectively improves the recognition accuracy of strong ionospheric scintillation while ensuring a high overall detection accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a strong ionospheric scintillation detection method based on a resampling-improved gradient boosting tree. Background Art

[0002] The ionosphere is an atmospheric layer about 60 to 1000 km above the ground, which is affected by ionization effects such as solar radiation and the geomagnetic field. Typical ionospheric scintillation phenomena can seriously affect radio waves such as global navigation satellite system (GNSS) signals. When GNSS signals pass through the scintillation ionosphere, amplitude and phase fluctuations, carrier Doppler jitter and other phenomena will occur, and even navigation interruption will occur. The Global Navigation Satellite System (GNSS) is widely used in various fields due to its all-weather, global coverage, high temporal and spatial resolution, low cost, and long-term stability. This makes ionospheric disturbances one of the key issues that must be solved in GNSS navigation and positioning technology. Minimizing the impact of the ionosphere on GNSS radio signals as much as possible to improve the accuracy of GNSS navigation and positioning is currently a very challenging and significant task.

[0003] Traditional ionospheric scintillation detection achieves simple classification by setting a hard threshold for the scintillation index, but the scintillation phenomenon manifests itself as continuous signal attenuation and phase jitter, making this detection method lack accuracy and flexibility in actual situations. The introduction of machine learning algorithms provides better performance for ionospheric scintillation recognition technology. This type of method can be based on the fluctuation of the ionospheric scintillation index and the comparison between different days, combined with the current satellite carrier-to-noise ratio, and use manual marking to exclude multipath interference and other information to obtain detection samples with higher accuracy and more flexible classification, thereby achieving a higher accuracy recognition effect. However, on the one hand, this type of method has a single sample ratio of scintillation of different intensities, lacks universality and applicability, and does not consider the impact of strong ionospheric scintillation samples, which account for a smaller proportion of the overall, on the accuracy of such scintillation recognition.

[0004] In recent years, machine learning methods have been introduced into ionospheric scintillation detection. Methods such as support vector machines and convolutional neural networks have achieved good results in detecting scintillation, providing efficient and flexible technical support for the development and measures of GNSS signal anti-ionospheric interference technology, and thus providing a major guarantee for the normal operation of satellite navigation systems. However, since strong scintillation occurs less frequently than other degrees of scintillation events, the proportion of strong ionospheric scintillation event samples in the data set used for model training is very small, and this imbalance will cause deviations in model training and affect the model prediction effect. Although the overall accuracy of the final model is relatively high, there is a certain amount of missed detection of strong ionospheric scintillation events. Summary of the invention

[0005] In view of the above shortcomings, the present invention provides a strong ionospheric scintillation detection method based on a resampling improved gradient boosting tree, wherein the method resamples the original training set containing less strong ionospheric scintillation using the SMOTE-ENN technology and then uses the XGBoost algorithm for classification, thereby realizing automatic detection of weak, medium and strong ionospheric scintillation events. Compared with the traditional method, the method can solve the problem of missed detection of strong scintillation caused by the imbalance of classification caused by the small number of strong ionospheric scintillation sample data in the training set, and the accuracy and efficiency of automatic identification of weak, medium and strong ionospheric scintillation events are relatively improved.

[0006] The strong ionospheric scintillation detection method based on the resampling improved gradient boosting tree has the following specific steps:

[0007] Step (1) extracts the mean of the correlation values ​​I and Q of the I and Q branches from the signal received by the GNSS receiver, and calculates I 2 , Q 2 And signal strength SI, SI 2 The mean of I, Q, and SI, and then the variance σ I , σ Q , σ SI And the covariance cov(I 2 ,Q 2 ) A total of 10 parameters are combined, the feature vectors of the undetected model are combined, the categories of weak, moderate and strong ionospheric scintillation are marked as 0, 1 and 2 respectively, and the feature vectors composed of the feature values ​​and the marked values ​​are combined into a sample matrix as the data set for model training and testing;

[0008] Step (2), in order to solve the problem that the number of strong ionospheric scintillation events in the training set is relatively small, the training samples obtained in step A are resampled, firstly, the minority class data in the sample is oversampled using the SMOTE oversampling technology, the minority class data includes strong ionospheric scintillation, that is, new samples are artificially synthesized according to the minority class samples and added to the training set, and then the majority class data in the training sample set is undersampled using the ENN undersampling technology, the majority class data includes weak and medium ionospheric scintillation, so as to obtain a new training set with a relatively balanced number of scintillation events;

[0009] Step (3) samples the resampled training set to obtain multiple subsets, and uses the XGBoost algorithm to train multiple decision trees with the same structure. Each round of decision trees learns the prediction residuals of the previous round and uses the second-order Taylor expansion to approximate the objective function. The regression tree that minimizes the objective function is found and the growth of the tree is controlled by setting the maximum depth of the tree and the sample weight threshold to prevent overfitting. The prediction scores corresponding to each tree are added together to obtain the final prediction result of the sample.

[0010] As a further improvement of the present invention, step (1) specifically comprises:

[0011] (1.1), the eigenvalue plays a decisive role in the classification model. At the end of the tracking process, the I, Q, branch signals and signal strength SI are extracted from the software GNSS receiver, and their mean and square mean are used as characteristic elements to reflect the ionospheric scintillation characteristics;

[0012] (1.2), considering the non-Gaussian noise characteristics of ionospheric scintillation, based on the above characteristic elements, another four variances and covariances are designed as characteristic values ​​reflecting ionospheric scintillation, expressed as follows:

[0013]

[0014]

[0015]

[0016]

[0017] where σ SI ,σ I ,σ Q are the variances of SI, I and Q respectively, cov(I 2 ,Q 2 ) is I 2 and Q 2 covariance, N is the number of samples used in each eigenvalue, and σ SI ,cov(I 2 ,Q 2 ),σ I ,σ Q The eigenvalue set is supplemented, and the final eigenvector is expressed as:

[0018] X={ , <q>,<I 2 >,<Q 2 >, <si>,<SI 2 >,σ SI ,cov(I 2 ,Q 2 ),σ I ,σ Q} (5)

[0019] The sampling frequencies of the I and Q channels are set to f = 1000 Hz. For each set of eigenvalues, the sliding window length T is set to 60 s, that is, each set of samples has N = T·f = 60000 initial observations. The sample frequency is set to 1 Hz, and the sliding window step size is 1 s;

[0020] (1.3) S4 is an index for measuring the ionospheric scintillation intensity. The larger it is, the higher the scintillation intensity. Calculate the S4 of each GNSS signal, and manually label the signals with 0 < S4 ≤ 0.2, 0.2 < S4 ≤ 0.6, and 0.6 ≤ S4 as scintillation events L = {0, 1, 2}, where 0, 1, and 2 correspond to weak, medium, and strong scintillation events respectively;

[0021] Label:

[0022] As a further improvement of the present invention, step (2) specifically includes:

[0023] (2.1) For the preprocessed training samples, perform linear interpolation between adjacent minority-class samples, and synthesize new minority-class samples and add them to the training samples. For each minority-class sample x m ∈X, m = 1,..., M, where X is the set of minority-class samples of the same kind, find k nearest neighbor samples K of the same kind, then select n samples from K according to the sampling rate n, and label them as y m,i ∈Y m , i = 1,..., n, m = 1,..., M, where Y m is the set of k nearest neighbor samples of the same kind for each x m in the set X of minority-class samples of the same kind. Finally, perform random linear interpolation between x m and y m,1 ,..., y m,n ;

[0024] The new minority-class samples are expressed as follows:

[0025] X new,m = x m + r * (Y m - x m ), m = 1,..., M (7)

[0026] where r is a random number from 0 to 1, and X new,m is the one including x m The vector of n new samples is established, so there are M*n new samples of the minority class;

[0027] (2.2) After obtaining a new training sample set from (2.1), the ENN method is used to undersample the majority class samples in the sample set. That is, for a majority class sample, if two or more different samples appear in the three nearest neighbor samples, the sample is deleted;

[0028] (2.3) Input the training set after deleting some majority class samples in step (2.2) into the XGBoost classifier for training.

[0029] As a further improvement of the present invention, step (3) specifically comprises:

[0030] (3.1) Initialize the first decision tree. Each round of decision trees learns the prediction residuals of the previous round. Obtain the training set from step (2.3), divide the training set into multiple subsets, and use each training subset to initialize some corresponding decision trees.

[0031] (3.2) In the process of tree generation, we look for a suitable regression tree that minimizes the objective function consisting of the loss function and the regularization term, which is composed of the difference between the predicted value and the true value and the regularization term, and is expressed as follows:

[0032]

[0033] Where i represents the i-th sample, n represents the total number of samples, and y i is the scoring result of the current tree for the sample; and refers to the predicted value of the previous tree, f t (x i ) is the new function of the current tree; Ω(f t ) is the corresponding regularization term; cons is the constant term;

[0034] (3.3) After using Taylor expansion on the objective function of (3.2) at f(x), Obj (t) Approximately:

[0035]

[0036] in T is the number of leaf nodes; ω j is the weight of the jth leaf node; γ and λ are regular penalty terms used to control the complexity of the tree. and the optimal solution Substituting into the above formula, the objective function of generating a new tree after optimization without considering the constant term cons is as follows:

[0037]

[0038] (3.4) Enumerate possible tree structures, calculate the objective function value in (3.3), and select the tree structure with the minimum value. Due to the large number of tree structures, a greedy algorithm is used to enumerate all different tree structures. Starting from a tree with a depth of 0, try to add a split node to each leaf node of the tree to achieve the maximum gain after the tree splits. For each feature, sort by feature value and use linear scanning to find the optimal split point of the feature. The gain expression is as follows:

[0039]

[0040] Among them, G L and G R is the first-order gradient value of the leaf nodes on the left and right sides after the tree bifurcates, H L and H R is the corresponding second-order gradient value. The first three terms are the left subtree score, the right subtree score, and the non-split score, respectively. γ is the complexity cost introduced by adding a new leaf node. The value of Gain determines whether the node of the decision tree continues to split.

[0041] (3.5) When the training is completed and k trees are obtained, the scores corresponding to each leaf node of each tree are added to obtain the sample prediction value, which is expressed as follows:

[0042]

[0043] F={f(x)=ω q(x) }(q:R m →T,ω∈R T ) (13)

[0044] Among them, f(x) is one of the decision trees, ω q(x) is the score of leaf node q;

[0045] (3.6) The sample prediction value is passed through the sigmoid function to realize the automatic classification of ionospheric scintillation. The final prediction category expression is as follows:

[0046]

[0047] Beneficial effects:

[0048] The frequency of strong ionospheric scintillation is less than that of weak / moderate scintillation. This imbalance phenomenon will have a great impact on the detection accuracy of strong ionospheric scintillation events. The present invention designs a strong ionospheric scintillation detection method based on a resampling improved gradient boosting tree. The method first uses the SMOTE-ENN technology to resample the training set to make the distribution of 0 / 1 / 2 class data more balanced. The XGBoost machine learning method is used in the training process of the automatic detection model to alleviate the overfitting problem in model training, optimize the model of decision tree training, and effectively improve the detection accuracy. The present invention effectively solves the imbalance problem of the training set caused by the natural and accidental occurrence of ionospheric scintillation events, and improves the detection accuracy of ionospheric scintillation, especially strong scintillation events. The present invention has universal significance for the interference detection of satellite signals (such as ionospheric scintillation, solar radio storms, deception), the design of advanced satellite receivers with higher accuracy, reliability and efficiency, and research in the fields of atmosphere and space weather research. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a strong ionospheric scintillation detection process based on the resampling improved gradient boosting tree;

[0050] Figure 2 This is the schematic diagram of the XGBoost algorithm. DETAILED DESCRIPTION

[0051] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. It should be noted that the words "front", "rear", "left", "right", "up" and "down" used in the following description refer to directions in the accompanying drawings, and the words "inside" and "outside" refer to directions toward or away from the geometric center of a specific component, respectively.

[0052] In conjunction with the accompanying drawings, the strong ionospheric scintillation detection method based on the resampling improved gradient boosting tree proposed in the present invention is described below. The overall structure is shown in the attached drawings. Figure 1 As shown, the following is a detailed description:

[0053] Step 1: Extract the mean of the correlation values ​​I and Q of the I and Q branches from the signal received by the GNSS receiver, and calculate the I 2 , Q 2 And signal strength SI, SI 2 The mean of I, Q, and SI, and then the variance σ I , σ Q , σ SI And the covariance cov(I 2 ,Q 2 ) A total of 10 parameters are combined to form the eigenvector of the undetected model, and the label categories of weak, medium and strong ionospheric scintillation are marked as 0, 1, and 2. The eigenvector composed of the eigenvalues ​​and the label value are combined into a sample matrix as the data set for model training and testing.

[0054] This step specifically includes:

[0055] (1.1) At the end of the tracking process, I, Q, branch signals and signal strength SI are extracted from the software GNSS receiver. Their mean and square mean are used as characteristic elements to reflect the ionospheric scintillation characteristics.

[0056] (1.2) Four other variances and covariances are designed as characteristic values ​​reflecting ionospheric scintillation: σ SI ,σ I ,σ Q are the variances of SI, I and Q respectively, cov(I 2 ,Q 2 ) is I 2 and Q 2 The covariance of .

[0057] in

[0058]

[0059]

[0060] The final feature vector is expressed as:

[0061] X={ , <q>,<I 2 >,<Q 2 >, <si>,<SI 2 >,σ SI ,cov(I 2 ,Q 2 ),σ I ,σ Q} (1)

[0062] The sampling frequencies of the I and Q channels are set to f = 1000 Hz. For each set of eigenvalues, the sliding window length is set to T = 60 s, that is, each set of samples has N = T·f = 60000 initial observations. The sample frequency is set to 1 Hz, and the sliding window step size is 1 s.

[0063] (1.3) S4 is an index to measure the ionospheric scintillation intensity. The larger it is, the higher the scintillation intensity. Calculate S4 for each GNSS signal, and manually label the signals with 0 < S4 ≤ 0.2, 0.2 < S4 ≤ 0.6, and 0.6 ≤ S4 as scintillation events L = {0, 1, 2}, corresponding to weak, medium, and strong scintillation events respectively.

[0064] Label:

[0065] Step 2: To solve the problem that the number of strong ionospheric scintillation events in the training set is relatively small, resample the training samples obtained in step A. First, use the SMOTE oversampling technique to oversample the minority class data in the samples. The minority class data includes strong ionospheric scintillation, that is, synthetic new samples are manually generated according to the minority class samples and added to the training set. Then use the ENN undersampling technique to undersample the majority class data in the training sample set. The majority class data includes weak and medium ionospheric scintillation, and a new training set with relatively balanced numbers of each scintillation event is obtained.

[0066] This step specifically includes:

[0067] (2.1) For the preprocessed training samples, perform linear interpolation between adjacent minority class samples. For each minority class sample x m ∈X (m = 1,..., M), find k nearest neighboring samples K of the same class, select n samples from K according to the sampling rate n, and label them as y m,i ∈Y m (i = 1,..., n, m = 1,..., M). Finally, perform random linear interpolation between x m and y m,1 ,..., y m,n . The new minority class samples are expressed as follows: X new,m =x m +r*(Y m -x m ), m = 1,..., M, r is a random number from 0 to 1, X new,m is included by x m Create a vector of n new samples.

[0068] (2.2) After obtaining a new training sample set from (2.1), the ENN method is used to undersample the majority class samples in the sample set. That is, for a majority class sample, if two or more different samples appear in the three neighboring samples, the sample is deleted.

[0069] (2.3) Input the training set after deleting some majority class samples in step (2.2) into the XGBoost classifier for training.

[0070] Step 3: Sample the resampled training set to obtain multiple subsets, using Figure 2 The XGBoost algorithm shown in the figure trains multiple decision trees with the same structure. Each round of decision trees learns the prediction residuals of the previous round and uses the second-order Taylor expansion to approximate the objective function, finds the regression tree that minimizes the objective function, and controls the growth of the tree by setting the maximum depth of the tree and the sample weight threshold to prevent overfitting. The prediction scores corresponding to each tree are added together to obtain the final prediction result of the sample.

[0071] (3.1) Initialize the first decision tree. Each round of decision trees learns the prediction residuals of the previous round. Obtain the training set from step (2.3), divide the training set into multiple subsets, and use each training subset to initialize some corresponding decision trees.

[0072] (3.2) In the process of tree generation, we look for a suitable regression tree that minimizes the objective function consisting of the loss function and the regularization term, which is composed of the difference between the predicted value and the true value and the regularization term, and is expressed as follows:

[0073]

[0074] Where i represents the i-th sample, n represents the total number of samples, and y i is the scoring result of the current tree for the sample; and Refers to the prediction value of the previous tree. t (x i ) is the new function of the current tree; Ω(f t ) is the corresponding regularization term; cons is the constant term.

[0075] (3.3) After using Taylor expansion on the objective function of (3.2) at f(x), Obj (t) It can be approximated as:

[0076]

[0077] in T is the number of leaf nodes; ω j is the weight of the jth leaf node; γ and λ are regular penalty terms used to control the complexity of the tree. and the optimal solution Substituting into the above formula, the objective function of generating a new tree after optimization without considering the constant term cons is as follows:

[0078]

[0079] (3.4) Enumerate possible tree structures, calculate the objective function value in (3.3), and select the tree structure with the minimum value. Due to the large number of tree structures, a greedy algorithm is used to enumerate all different tree structures. Starting from a tree with a depth of 0, each leaf node of the tree is tried to add a split node to achieve the maximum gain after the tree splits. For each feature, sort by feature value and use linear scanning to find the optimal split point of the feature. The gain expression is as follows:

[0080]

[0081] Among them, G L and G R is the first-order gradient value of the leaf nodes on the left and right sides after the tree bifurcates, H L and H R is the corresponding second-order gradient value. The first three terms are the left subtree score, the right subtree score, and the non-split score, respectively, and γ is the complexity cost introduced by adding a new leaf node. The value of Gain determines whether the node of the decision tree continues to split.

[0082] (3.5) When the training is completed and k trees are obtained, the scores corresponding to each leaf node of each tree are added to obtain the sample prediction value, which is expressed as follows:

[0083]

[0084] F={f(x)=ω q(x) }(q:R m →T,ω∈R T ) (8)

[0085] Among them, f(x) is one of the decision trees, ω q(x) is the score of leaf node q.

[0086] (3.6) The sample prediction value is passed through the sigmoid function to realize the automatic classification of ionospheric scintillation. For the three-classification problem of the present invention, the final prediction category expression is as follows:

[0087]

[0088] Finally, it should be noted that the above description is only a preferred example of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.< / si> < / q> < / si> < / q>

Claims

1. A strong ionospheric scintillation detection method based on a resampling improved gradient boosting tree, characterized in that: The specific steps are as follows: Step (1) extracts the mean of the correlation values ​​I and Q of the I and Q branches from the signal received by the GNSS receiver, and calculates I 2 , Q 2 And signal strength SI, SI 2 The mean of I, Q, and SI, and then the variance σ I , σ Q , σ SI And the covariance cov(I 2 ,Q 2 ) A total of 10 parameters are combined, the feature vectors of the undetected model are combined, the categories of weak, moderate and strong ionospheric scintillation are marked as 0, 1 and 2 respectively, and the feature vectors composed of the feature values ​​and the marked values ​​are combined into a sample matrix as the data set for model training and testing; Step (1) specifically includes: (1.1) The eigenvalue plays a decisive role in the classification model. At the end of the tracking process, the I, Q, branch signals, and signal strength SI are extracted from the software GNSS receiver, and their mean values and squared mean values are used as feature elements to reflect the ionospheric scintillation characteristics; (1.2) Considering the non-Gaussian noise characteristics of ionospheric scintillation, based on the above feature elements, another 4 variances and covariances are designed as eigenvalues reflecting ionospheric scintillation, and the expression formulas are as follows: where σ SI ,σ I ,σ Q are the variances of SI, I and Q respectively, cov(I 2 ,Q 2 ) is I 2 and Q 2 covariance, N is the number of samples used in each eigenvalue, and σ SI ,cov(I 2 ,Q 2 ),σ I ,σ Q The eigenvalue set is supplemented, and the final eigenvector is expressed as: X={ , <q>,<I 2 >,<Q 2 >, <si>, <SI 2 >,s SI ,cov(I 2 ,Q 2 ),s I ,s Q } (5)< / si> < / q> The sampling frequencies of the I and Q channels are set to f = 1000 Hz. For each group of eigenvalues, the sliding window length T = 60 s is set, that is, each group of samples has N = T·f = 60000 initial observation values, the sample frequency is set to 1 Hz, and the sliding window step size is 1 s; (1.3) S4 is an index to measure the ionospheric scintillation intensity. The larger it is, the higher the scintillation intensity. Calculate S4 of each GNSS signal, and manually label the signals with 0 < S4 ≤ 0.2, 0.2 < S4 ≤ 0.6, and 0.6 < S4 as scintillation events L = {0, 1, 2}, where 0, 1, and 2 correspond to weak, medium, and strong scintillation events respectively; Label: Step (2), to solve the problem that the number of strong ionospheric scintillation events in the training set is relatively small, resample the training samples obtained in step A. First, use the SMOTE oversampling technique to oversample the minority class data in the samples. The minority class data includes strong ionospheric scintillation, that is, new samples are artificially synthesized according to the minority class samples and added to the training set. Then, use the ENN undersampling technique to undersample the majority class data in the training sample set. The majority class data includes weak and medium ionospheric scintillation, and a new training set with relatively balanced numbers of each scintillation event is obtained; Step (2) specifically includes: (2.1) For the preprocessed training samples, linear interpolation is performed between adjacent minority class samples to synthesize new minority class samples and add them to the training samples. For each minority class sample x m ∈X, m=1,...,M, where X is a set of a few similar samples, find the k nearest similar adjacent samples K, and then select n samples from K according to the sampling rate n, and mark them as y m,i ∈Y m , i=1,...,n,m=1,...,M, where Y m is a small number of similar samples in the set X where each x m The set of k nearest similar samples, finally in x m and m,1 ,...,y m,n Random linear interpolation is performed between ; The new minority class samples are expressed as follows: X new,m =x m +r*(Y m -x m ),m=1,...,M (7) Among them, r is a random number between 0 and 1, X new,m is included by x m The vector of n new samples is established, so there are M*n new samples of the minority class; (2.2) After obtaining the new training sample set from (2.1), use the ENN method to undersample the majority class samples in this sample set. That is, for a certain majority class sample, if two or more different samples appear among the three nearest neighbor samples, then delete this sample; (2.3) Input the training set after deleting some majority class samples in step (2.2) into the XGBoost classifier for training; Step (3), sample the resampled training set to obtain multiple subsets, and use the XGBoost algorithm to train multiple decision trees with the same structure. Each round of decision tree learns the prediction residuals of the previous round and approximates the objective function using the second-order Taylor expansion. Find the regression tree that minimizes the objective function and control the growth of the tree by setting the maximum depth of the tree and the sample weight threshold to prevent overfitting. Add the prediction scores corresponding to each tree to obtain the final prediction result of the sample; Step (3) specifically includes: (3.1) Initialize the first decision tree. Each round of decision tree learns the prediction residuals of the previous round. Obtain the training set from step (2.3), divide the training set into multiple subsets, and initialize some corresponding decision trees using each training subset; (3.2) In the process of tree generation, we look for a suitable regression tree that minimizes the objective function consisting of the loss function and the regularization term, which is composed of the difference between the predicted value and the true value and the regularization term, and is expressed as follows: Where i represents the i-th sample, n represents the total number of samples, and y i is the scoring result of the current tree for the sample; and refers to the predicted value of the previous tree, f t (x i ) is the new function of the current tree; Ω(f t ) is the corresponding regularization term; cons is the constant term; (3.3) After using Taylor expansion on the objective function of (3.2) at f(x), Obj (t) Approximately: in T is the number of leaf nodes; ω j is the weight of the jth leaf node; γ and λ are regular penalty terms used to control the complexity of the tree. and the optimal solution Substituting into formula (9), the objective function of generating a new tree after optimization without considering the constant term cons is as follows: (3.4) Enumerate possible tree structures, calculate the objective function value in (3.3), and select the tree structure with the minimum value. Due to the large number of tree structures, a greedy algorithm is used to enumerate all different tree structures. Starting from a tree with a depth of 0, try to add a split node to each leaf node of the tree to achieve the maximum gain after the tree splits. For each feature, sort by feature value and use linear scanning to find the optimal split point of the feature. The gain expression is as follows: Among them, G L and G R is the first-order gradient value of the leaf nodes on the left and right sides after the tree bifurcates, H L and H R is the corresponding second-order gradient value, where the first three terms are the left subtree score, the right subtree score, and the non-split score, respectively. γ is the complexity cost introduced by adding a new leaf node. The value of Gain determines whether the node of the decision tree continues to split. (3.5) When the training is completed and k trees are obtained, the scores corresponding to each leaf node of each tree are added to obtain the sample prediction value, which is expressed as follows: Among them, f(x) is one of the decision trees, ω q(x) is the score of leaf node q; (3.6) The sample prediction value is passed through the sigmoid function to realize the automatic classification of ionospheric scintillation. The final prediction category expression is as follows:

Citation Information

Patent Citations

  • An ionospheric phase scintillation detection method based on nonlinear SVM algorithm

    CN109508730A

  • UHF frequency band ionospheric scintillation event forecasting method based on TEC integral quantity and seasonal variation coefficient

    CN114089443A

  • System and Method for GNSS Ambiguity Resolution

    US20190302274A1