Classification method for exoplanet light curve signals based on data enhancement and ensemble learning

Abstract

The application discloses an exoplanet light variation signal classification method based on data enhancement and integrated learning, which comprises the following steps: splicing light variation signal data of different observation seasons of the same target, and pre-processing the data; resampling the data to a fixed length; dividing the resampled data set into a training set, a test set and a validation set; performing data enhancement on the training set by over-sampling and under-sampling respectively to obtain sub-training set 1 and sub-training set 2; building a one-dimensional CNN network model for the two training sets respectively, and optimizing the network by adjusting the number of hidden layers, the number of neurons, the activation function and the model optimizer; obtaining the best classification model as base classifier 1 and base classifier 2 by adopting N-fold cross-validation respectively; combining the base classifiers by adopting a weighted voting method to obtain a light variation signal classifier, and obtaining a classification result according to the signal classifier. The application solves the problem of unbalanced exoplanet light variation signal categories, and enables the model to obtain more features to improve the generalization ability.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent application of deep learning algorithm, and relates to an exoplanet light variation signal classification method based on data enhancement and ensemble learning. BACKGROUND

[0002] With the development of telescope construction technology and scientific and technological level, the era of time-domain astronomy taking dynamic astronomical events as research objects is opened. Astronomers fit light variation signals through time series data of target regions to study the changes of target celestial bodies over time. Massive data requires that the data processing pipeline has the ability of fast automatic response, automatic observation and real-time processing of temporary source trigger signals. The data-driven deep learning method needs a large number of samples to obtain ideal results, however, in the data obtained by the exoplanet search method, the planet target sample is difficult to obtain.

[0003] In multi-class imbalance classification, problems such as class overlap, noise and multiple minority classes reduce the ability of the classifier, which can be improved from the data preprocessing level and the algorithm level. On the data level, the original data set can be modified through data preprocessing, including undersampling, oversampling, hybrid sampling and other resampling methods to optimize the sample space; on the algorithm level, the classification algorithm can be improved or integrated to make it more suitable for the requirements of unbalanced data classification. SUMMARY

[0004] In view of the above prior art, the technical problem to be solved by the application is to provide an exoplanet light variation signal classification method based on data enhancement and ensemble learning, which integrates data level methods and algorithm level methods in a unified framework, solves the problem of lack of positive samples in astronomical time series data, and meets the application requirements of deep learning in the field of astronomical time series under multi-class imbalance.

[0005] To solve the above technical problem, an exoplanet light variation signal classification method based on data enhancement and ensemble learning inputs the exoplanet light variation signal to be classified into a pre-constructed light variation signal classifier, and outputs a signal classification result.

[0006] The construction method of the light variation signal classifier comprises:

[0007] S1: splicing light variation signal data of different observation seasons of the same target, and pre-processing the spliced data;

[0008] S2: resampling the pre-processed light variation signal data to a fixed length;

[0009] S3: dividing the resampled data set into a training set, a test set and a validation set;

[0010] S4: Use oversampling and undersampling respectively to perform data augmentation on the training set, and obtain sub-training set 1 and sub-training set 2;

[0011] S5: Build a one-dimensional CNN network model for sub-training set 1 and sub-training set 2 respectively, and optimize the network by adjusting the number of hidden layers, the number of neurons, the activation function, and the model optimizer;

[0012] S6: Use N-fold cross-validation to obtain the best classification models as base classifier 1 and base classifier 2 respectively;

[0013] S7: Combine the base classifiers using the weighted voting method to obtain the optical change signal classifier.

[0014] Furthermore, the preprocessing of the spliced data includes missing value processing, outlier processing, moving smoothing processing, and normalization processing.

[0015] Furthermore, the resampling of the preprocessed optical change signal data to a fixed length includes: resampling the preprocessed signal into n points. Signals with a data length less than n are complemented by linear interpolation, and signals with a length greater than n are downsampled by taking the mean within a sliding window.

[0016] Furthermore, the use of oversampling to perform data augmentation on the training set includes:

[0017] Step S4.1: Initialize the minority class sample space S min,n ={x1,x2,…,x n}, the majority class sample space S max ;

[0018] Step S4.2: Use the KNN algorithm to determine the K nearest samples of the sample x n , denoted as X nk ={x n1 ,x n2 ,…,x nk};

[0019] Step S4.3: Randomly select m samples (m < k) from X nk , and perform random linear interpolation with the sample x n respectively to synthesize new minority class samples close to the original data distribution. The new sample distribution satisfies:

[0020]

[0021] Combine the oversampled minority class samples with the original majority class samples to obtain the new sub-training set 1.

[0022] Furthermore, the use of undersampling to perform data augmentation on the training set includes:

[0023] Step S4.a: initialize the training set D1={x1,x2,…,x n}, sample point x i ={x i1 ,x i2 ,…,x im}, majority sample space S max , the number of clusters is c, and the number of under-sampled samples is N;

[0024] Step S4.b: calculate the weight of different dimension data of the sample:

[0025]

[0026] where x ir is the size of the rth component value in the ith data, and the Euclidean distance between the samples in S max after dimension weighting is calculated:

[0027]

[0028] where dist w (x i ,x j ) is the Euclidean distance calculated in m dimensions of two samples, and the sum of the distances between each pair of sample points:

[0029]

[0030] Step S4.c: calculate the sample with the smallest distance sum as the first sample cluster center, and take this sample as the first cluster center. Calculate the distance between any sample and the first cluster center, and take the sample with the maximum distance as the second cluster center. In this way, the cth cluster center is obtained.

[0031] Step S4.d: select N-c samples with the smallest distance to the cluster center, and select samples for each cluster. Finally, a sample set containing N samples is obtained.

[0032] Step S4.e: obtain the under-sampled majority sample and the original minority sample combined sub-training set 2.

[0033] Further, N-fold cross-validation uses Macro-F1 as the evaluation index of the classification task, and selects the optimal model as the base separator according to the Macro-F1 evaluation result. Macro-F1 satisfies:

[0034]

[0035] Where:

[0036]

[0037]

[0038] Wherein: C is the number of categories, precision i is the precision of the i-th category, recall i is the recall of the i-th category, C ii refers to samples with true labels i and classification results i, C ·i and C i· respectively refer to all samples classified as i and all samples with true labels as i.

[0039] Further, the light variation signal classifier obtained by combining the base classifiers by using the weighted voting method comprises:

[0040]

[0041] Wherein, w i is the weight of the base classifier i, and the macro-F1 value is selected as the weight w i of each model, and h i (x) is the classification result of the base classifier i.

[0042] The beneficial effects of the present application are:

[0043] The present application adopts a data enhancement method, respectively based on SMOTE algorithm oversampling to expand the minority class, based on clustering undersampling to reduce the majority class, and respectively obtains a class-balanced sub-training set. The class resampling method reduces the variance of the base classifier, improves the generalization error, and reduces the error caused by random fluctuations in the original training set.

[0044] The present application adopts an ensemble learning method, which provides a solution for the multi-class imbalance problem of exoplanet signals by combining multiple learners. By combining multiple simple models to obtain a more optimal combined model, the resampling technology and the ensemble learning method adapt to the imbalance distribution of the classes at the algorithm level. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 It is a multi-class imbalance light variation data classification method of the present application.

[0046] Figure 2 It is an algorithm architecture diagram of the present application.

[0047] Figure 3 It is a convolutional neural network structure diagram of the base classifier of the present application. DETAILED DESCRIPTION

[0048] The present application will be further described below in conjunction with the drawings and examples of the specification.

[0049] The present application aims to solve the problem of small positive sample data in the field of exoplanet exploration by the transit method, and the reduction of the ability of the classifier due to the multi-class unbalanced data set. It includes: data preprocessing of the light variation signal obtained by the transit method for exoplanet exploration, and dividing the data set; over-sampling of the minority samples and under-sampling of the majority samples are carried out on the training set respectively to obtain sub-training sets; neural network models are built based on different sub-training sets; the optimal classification model is selected by cross-validation; and multi-model weighted fusion is carried out according to the evaluation results. The method expands the signal samples and balances the number of categories through data enhancement, and solves the problem of insufficient samples of exoplanet light variation signals from the data level. Through ensemble learning, the model can obtain more features to improve the generalization ability from the algorithm level.

[0050] The present application comprises the following steps:

[0051] The light variation signal of the exoplanet to be classified is input into the pre-trained light variation signal classifier, and the signal classification result is output;

[0052] The training method of the light variation signal classifier comprises:

[0053] S1: splicing light variation data of different observation seasons of the same target;

[0054] S2: preprocessing the spliced light variation signal and resampling the samples to a fixed length;

[0055] S3: dividing the data set into a training set, a test set and a validation set;

[0056] S4: data enhancement of the training set by over-sampling and under-sampling;

[0057] S5: based on the sub-training sets after different data enhancement strategies, one-dimensional CNN network models are built respectively;

[0058] S6: N-fold cross-validation to obtain the best classification model;

[0059] S7: integrating the models to comprehensively utilize the advantages of multiple models.

[0060] Step S2 preprocessing specifically includes missing value processing, outlier processing, moving smoothing, normalization, and the specific process is:

[0061] Step S2.1: the missing value processing deletes all NaN value data points;

[0062] Step S2.2: the outlier processing deletes extreme large values and extreme small values far from the local general level of the sequence.

[0063] Step S2.3: For the moving smoothing process, the Savitzky-Golay method is used to perform k-order polynomial fitting on the data points within a window of a certain length to obtain the fitting result.

[0064] Step S2.4: The normalization is a linear transformation of the original data, which scales the data to the interval [0,1]. The transformation formula is:

[0065]

[0066] where max is the maximum value of the sample and min is the minimum value of the sample.

[0067] Step S2.5: For the resampling, the preprocessed signal is resampled into n points. Signals with a data length less than n are complemented by linear interpolation, and signals with a length greater than n are downsampled by taking the mean within a sliding window.

[0068] The specific process of Step S3 is as follows:

[0069] Step S3.1: The data set is divided into a training set and a test set in a ratio of 3:1.

[0070] Step S3.2: The training set is further divided into a cross-training set and a cross-validation set in a ratio of 4:1.

[0071] The specific process of Step S4 is as follows:

[0072] The SMOTE oversampling algorithm is used to synthesize the minority class for the original training set to obtain the sub-training set 1.

[0073] Step S4.1: Initialize the minority class sample space S min,n ={x1,x2,…,x n}, the majority class sample space S max ;

[0074] Step S4.2: Use the KNN algorithm to determine the K nearest samples of the sample x n , denoted as X nk ={x n1 ,x n2 ,…,x nk};

[0075] Step S4.3: Randomly select m samples (m < k) from X nk , and perform random linear interpolation with the sample x n respectively to synthesize new minority class samples close to the original data distribution, thereby achieving the balance of the data set. The calculation formula for the distribution of the new samples is as follows:

[0076]

[0077] The oversampled minority class samples are combined with the original majority class samples to obtain a new sub-training set 1.

[0078] The clustering-based undersampling is used on the original training set to obtain a training set 2.

[0079] Step S4.4: initialization of the training set D1={x1,x2,…,x n}, sample points x i ={x i1 ,x i2 ,…,x im}, a majority class sample space S max , the number of clusters c, and the number of undersampled samples N;

[0080] Step S4.5: calculation of the weight of the different dimension data of the sample:

[0081]

[0082] where x ir is the size of the r-th component value in the i-th data, and the Euclidean distance between the samples in S max after dimension weighting is calculated:

[0083]

[0084] where dist w (x i ,x j ) is the Euclidean distance calculated in m dimensions of two samples, and the sum of the distances between two sample points:

[0085]

[0086] Step S4.6: the sample with the smallest sum of distances is regarded as the first sample cluster center, and this sample is regarded as the first cluster center. The distance between any sample and the first cluster center is calculated, and the sample with the maximum distance is taken as the second cluster center. In this way, the c-th cluster center is taken;

[0087] Step S4.7: select N-c samples with the smallest distance to the cluster center, and select samples for each cluster, and finally obtain a sample set containing N samples;

[0088] Step S4.8: the obtained undersampled majority class samples are combined with the original minority class samples to obtain a new sub-training set 2.

[0089] The specific process of step S5 is:

[0090] Step S5.1: based on the sub-training set 1 and the sub-training set 2, a one-dimensional convolutional neural network model is constructed respectively;

[0091] Step S5.2: optimizing the network by adjusting the number of hidden layers, the number of neurons, the activation function and the model optimizer.

[0092] The specific process of step S6 is:

[0093] Step S6.1: using 5-fold cross-validation, two training sets are tested on 5 different combinations, called one cross-validation;

[0094] Step S6.2: in each cross-validation, Macro-F1 is used as the evaluation index of the classification task, and the formula is as follows:

[0095]

[0096] Wherein:

[0097]

[0098]

[0099] Wherein: C is the number of categories, precision i is the precision of the i-th category, recall i is the recall of the i-th category, C ii means that the true label is i and the classification result is also i, C ′i and C i· respectively refer to all samples classified as i and all samples with true label i.

[0100] The specific process of step S7 is:

[0101] Step S7.1: selecting the optimal model as the base learner according to the Macro-F1 evaluation result;

[0102] Step S7.2: combining individual classifiers according to the weighted voting method, and the formula is as follows:

[0103]

[0104] Wherein, w i is the weight of the i-th learner, and h i (x) is the classification result of the base classifier.

[0105] The following embodiments are given in conjunction with the drawings:

[0106] First part: preprocessing of light variation signal.

[0107] Splicing light variation data of different observation seasons of the same target;

[0108] Missing value handling will delete all data points with NaN values;

[0109] Outlier handling will delete extremely large and extremely small values that are far from the local general level of the sequence.

[0110] Moving smoothing processing uses Savitzky-Golay to fit the data points within a window of a certain length with a k-order polynomial to obtain the fitted result.

[0111] Normalization processing is a linear transformation of the original data, which scales the data to the interval [0,1]. The transformation formula is:

[0112]

[0113] where max is the maximum value of the sample and min is the minimum value of the sample.

[0114] Resample the samples to a fixed length. Resample the preprocessed signal into n points. Signals with a data length less than n are completed by linear interpolation, and signals with a length greater than n are downsampled by taking the mean within a sliding window.

[0115] Divide the dataset into a training set, a test set, and a validation set respectively.

[0116] Divide the dataset into a training set and a test set in a ratio of 3:1;

[0117] Further divide the training set into a cross-training set and a cross-validation set in a ratio of 4:1.

[0118] Part Two: Data augmentation of optical variation signals.

[0119] Use the SMOTE oversampling algorithm to synthesize the minority class for the original training set to obtain the sub-training set 1.

[0120] Initialize the minority class sample space S min,n ={x1,x2,…,x<> n}, the majority class sample space S max ;

[0121] Use the KNN algorithm to determine the K nearest samples of the sample x n and denote them as X nk ={x n1 ,x n2 ,…,x nk}; <000>

[0122] Randomly select m samples (m < k) from X nk and separately combine them with the sample x nRandom linear interpolation is performed to synthesize new minority class samples that approximate the original data distribution, thereby achieving dataset balance. The formula for calculating the distribution of the new samples is as follows:

[0123]

[0124] Cluster-based undersampling was applied to the original training set to obtain training set 2.

[0125] Initialize the training set D1 = {x1, x2, ..., x} n}, sample point x i ={x i1 ,x i2 ,…,x im}, majority class sample space S max The number of clusters is c, and the number of undersampled samples is N;

[0126] Calculate the weights of the data in different dimensions of the sample:

[0127]

[0128] Where, x ir Let S be the value of the r-th component in the i-th data, and calculate S. max Euclidean distance between samples in the sample set after dimensional weighting:

[0129]

[0130] Among them, dist w (x i ,w j The distance is the Euclidean distance between two samples in the m-dimensional space, and the sum of the distances between any two sample points.

[0131]

[0132] The sample with the minimum sum of distances is considered the first cluster center. Calculate the distance between any sample and the first cluster center, and take the sample with the maximum distance as the second cluster center, and so on, until the c-th cluster center is selected.

[0133] Select the Nc samples with the smallest distance from the cluster center, and select samples from each cluster. From one sample, we get a sample set containing N samples.

[0134] The resulting undersampled majority class samples are combined with the original minority class samples to form a new sub-training set 2.

[0135] Part 3: Feature extraction of optically variable signals.

[0136] Based on the sub-training set after different data enhancement strategies, a one-dimensional CNN network model is built;

[0137] By adjusting the number of hidden layers, the number of neurons, the activation function and the model optimizer, the network is optimized.

[0138] Two training sets are tested on five different combinations, called one cross-validation;

[0139] In each cross-validation, Macro-F1 is used as the evaluation index of the classification task, and the formula is as follows:

[0140]

[0141] Among them:

[0142]

[0143]

[0144] According to the Macro-F1 evaluation result, the optimal model is selected as the base classifier.

[0145] The fourth part: integrated learning of light variation signal.

[0146]

[0147] Among them, w i is the weight of the ith learner, and h i (x) is the classification result of the base classifier.

[0148] For the problems of insufficient samples and multi-class imbalance in exoplanet light variation signal recognition, the invention is improved from data enhancement and integrated learning. In the aspect of data enhancement, smote oversampling and clustering-based undersampling are combined to balance the data set; in the aspect of integrated learning, multiple classifiers are weighted and voted to improve the classification accuracy.

[0149] It should be understood that although the present specification is described in terms of embodiments, each embodiment does not contain only one independent technical solution, and the description manner of the specification is only for clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be combined to form other embodiments which can be understood by those skilled in the art. The classification method based on data enhancement and integrated learning makes up for the lack of positive samples of light variation data and the multi-class imbalance problem. The integrated classifier constructed has good generalization performance, effectively solves the overfitting and information loss problem caused by unbalanced data set.

Claims

1. A method for classifying exoplanet light variation signals based on data augmentation and ensemble learning, characterized in that: The light variation signals of the exoplanets to be classified are input into a pre-constructed light variation signal classifier, and the signal classification results are output. The method for constructing the optical signal classifier includes: S1: Stitch together light variation signal data of the same target from different observation seasons, and preprocess the stitched data; S2: Resample the preprocessed optical variable signal data to a fixed length; S3: Divide the resampled dataset into a training set, a test set, and a validation set; S4: Oversampling and undersampling are used to augment the training set to obtain sub-training set 1 and sub-training set 2 respectively; Step S4.1: Initialize the minority class sample space majority class sample space ; Step S4.2: Use the KNN algorithm to determine the samples The K nearest neighbors are denoted as ; Step S4.3: Randomly select m samples Compare them with the samples respectively Perform random linear interpolation to synthesize new minority class samples that approximate the original data distribution. The new sample distribution satisfies: ； The oversampled minority class samples are combined with the original majority class samples to obtain a new sub-training set 1; Step S4.4: Initialize the training set Sample points majority class sample space The number of clusters is c, and the number of undersampled samples is N; Step S4.5: Calculate the weights of different dimensions of the sample data: ； in, Let r be the value of the r-th component in the i-th data, and calculate... Euclidean distance between samples in the sample set after dimensional weighting: ； in, It is the Euclidean distance between two samples in m dimensions, the sum of the distances between any two sample points: ； Step S4.6: The sample with the minimum sum of distances is considered the first sample cluster center. This sample is considered the first cluster center. The distance between any sample and the first cluster center is calculated. The sample with the maximum distance is taken as the second cluster center, and so on, until the cth cluster center is taken. Step S4.7: Select the cluster with the smallest distance from the cluster center. Samples, selected from each cluster From one sample, we get a sample set containing N samples. Step S4.8: The resulting sub-training set 2 is a combination of the undersampled majority class samples and the original minority class samples; S5: Build one-dimensional CNN network models for sub-training set 1 and sub-training set 2 respectively, and optimize the network by adjusting the number of hidden layers, the number of neurons, the activation function and the model optimizer; S6: Use N-fold cross-validation to obtain the best classification model as base classifier 1 and base classifier 2 respectively; S7: A weighted voting method is used to combine base classifiers to obtain a classifier for optical variable signals.

2. The method for classifying exoplanet light variation signals based on data augmentation and ensemble learning according to claim 1, characterized in that: The preprocessing of the spliced ​​data includes missing value processing, outlier processing, moving smoothing, and normalization.

3. The method for classifying exoplanet light variation signals based on data augmentation and ensemble learning according to claim 1, characterized in that: The process of resampling the preprocessed optical signal data to a fixed length includes: resampling the preprocessed signal into n points; for signals with a data length less than n, linear interpolation is used for completion, and for signals with a length greater than n, the mean is taken within a sliding window for downsampling.

4. The method for classifying exoplanet light variation signals based on data augmentation and ensemble learning according to claim 1, characterized in that: N-fold cross-validation uses Macro-F1 as the evaluation metric for the classification task. The optimal model is selected as the basis separator based on the Macro-F1 evaluation results. Macro-F1 satisfies the following conditions: ； in: ； ； Where: C represents the number of categories. Let be the precision for the i-th category. Let be the recall rate for the i-th category. , , This refers to samples whose true label is i and whose classification result is also i. and These refer to all samples classified as i and all samples with the true label i, respectively.

5. The method for classifying exoplanet light variation signals based on data augmentation and ensemble learning according to claim 4, characterized in that: The method of combining base classifiers using weighted voting to obtain a classifier for optical variable signals includes: ； in, The macro-F1 score is used as the weight for each model, representing the weights of the base classifier i. , The classification result of base classifier i.

Images