An exoplanet light variation signal classification method based on transfer learning

Abstract

The application discloses an exoplanet light variation signal classification method based on transfer learning, acquires light variation signals of two telescopes for detecting exoplanets by using the transit method, and respectively uses the light variation signals as a pre-training model using dataset 1 and a target model using dataset 2; the acquired data is preprocessed; the two dataset data are respectively divided into a training set and a test set; a convolutional neural network model is constructed by using the dataset 1 to perform training, an optimal time series data classification model is established to perform time series classification; the pre-training model is saved; the dataset 2 is used to adjust the model, and a model suitable for the dataset 2 is obtained. The application avoids the limitation that a model cannot be learned under the condition of a small sample, and accelerates intelligent data exoplanet detection.

Description

Technical Field

[0001] This invention belongs to the field of intelligent applications of deep learning algorithms, and relates to a method for classifying exoplanet light variation signals based on transfer learning. Background Technology

[0002] Currently, in the field of astronomy, observational equipment and technologies are developing rapidly, with telescopes continuously improving data acquisition rates and data volumes constantly increasing. Astronomers urgently need artificial intelligence methods to analyze massive amounts of data and extract knowledge from them. With the implementation of large-field-of-view rapid sky surveys, astronomy has entered the era of time-domain astronomy. In time-series data analysis, exoplanet search has become one of the hottest research directions in astronomy in recent years. Among the methods for searching exoplanets, the transit method has found the most exoplanets to date. Detecting periodic transit events based on set thresholds cannot rely entirely on manual verification with large datasets and requires automated assistance.

[0003] Light variation signal time series data are curves showing the change in celestial object brightness over time. In time-domain astronomy, data acquisition is measured in seconds, and data duplication and invalidity increase the negative samples in the dataset. Actual transit phenomena can have various possibilities, potentially manifesting differently in the light variation sequence, requiring the model to be robust to noise. Data acquired by a single telescope contains very few positive samples, making it difficult to train a network model to achieve good results.

[0004] Transfer learning no longer assumes that training and test data must be independent and identically distributed. Therefore, the target model does not need to be trained from scratch, reducing the requirements for training data and accelerating the training speed. In the area of ​​exoplanet detection, there are completed sky survey projects, such as the Kepler Space Telescope, which has become a major data provider. The verified exoplanet light variation data from these completed survey projects can be used as a large-sample, labeled dataset. Transfer learning can be employed to obtain a pre-trained model using this large dataset, filling the gap in domestic exoplanet detection and accelerating model training and data processing pipeline development. Summary of the Invention

[0005] In view of the above-mentioned prior art, the technical problem to be solved by the present invention is to provide a method for classifying exoplanet light variation signals based on transfer learning, which solves the problems of insufficient labeled data and small number of positive samples in astronomical time series data, and meets the application requirements of deep learning in the field of astronomical time series under small sample conditions.

[0006] To address the aforementioned technical problems, this invention provides a method for classifying exoplanet light variation signals based on transfer learning. The method inputs the exoplanet light variation signals to be classified into a pre-trained light variation signal classification model and outputs the signal classification results.

[0007] The optical signal classification model employs a one-dimensional convolutional neural network, including alternating convolutional and pooling layers, fully connected layers, and softmax layers. The training method includes:

[0008] S1: Select two telescopes that use the transit method to observe exoplanets. Obtain light variation signals from different telescopes through a catalog of verified exoplanets. Select the telescope data with larger data volume as large sample dataset 1 and the telescope data with smaller data volume as small sample dataset 2.

[0009] S2: In two datasets, samples from multiple observation quarters of the same target are spliced ​​together. The spliced ​​light variation signal from the two datasets is then processed for missing values, outliers, moving smoothing, and normalization.

[0010] S3: Resample each optical variable signal sample to a fixed length;

[0011] S4: Divide both dataset 1 and dataset 2 into training set and validation set;

[0012] S5: Use dataset 1 to train the constructed one-dimensional convolutional neural network and save the optimal light-varying signal classification model as the source domain pre-training model;

[0013] S6: Continue training the pre-trained model using dataset 2, transfer the source domain model to the target domain, and adjust to obtain the optimal light variation signal classification model for the small sample dataset 2.

[0014] Furthermore, the outlier processing involves removing outliers that are more than one standard deviation above the median σ1 and less than the median σ2.

[0015] Furthermore, the motion smoothing process includes:

[0016] A weighted moving average smoothing filter, Savitzky-Golay, is used to perform k-order polynomial fitting on the data points within the window to remove background noise.

[0017]

[0018] in, The smoothing coefficient is obtained by fitting a polynomial using the least squares method.

[0019] Furthermore, the normalization process includes:

[0020] Perform Min-Max normalization on the samples:

[0021]

[0022] Where max is the maximum value of the optical variation data and min is the minimum value of the optical variation data.

[0023] Furthermore, S3's resampling of each optically variable signal sample to a fixed length includes:

[0024] The discrete Fourier method is used to resample the spliced ​​signal into n points, with the original sequence interval changed from d. x Become Specifically:

[0025] X = dft(x)

[0026] y = idft(X, n)

[0027] Where x represents the original sequence, X is the result of the discrete Fourier transform of x, n is the number of points in the target sequence, and y is the result of the discrete inverse Fourier transform of X.

[0028] Furthermore, training the constructed one-dimensional convolutional neural network using dataset 1 includes:

[0029] S1: Initialize the parameters of the one-dimensional convolutional neural network;

[0030] S2: After the network forward propagation passes through alternating convolutional and pooling layers, fully connected layers, and softmax layers, the network error is calculated, and it is determined whether the error meets the requirements. If not, the network is updated by adjusting the network combination and hyperparameters until the network error or the number of iterations reaches the set requirements.

[0031] S3: Determine if the accuracy of the validation set meets the requirements. If it does, save the network model; otherwise, reconfigure the network structure and hyperparameters, and return to S2.

[0032] Furthermore, the step of using dataset 2 to continue training the pre-trained model and transferring the source domain model to the target domain includes:

[0033] S(1): Transfer the network structure and parameters of the pre-trained model to the new model, and use the parameters of the pre-trained model as the initial values ​​of the parameters of the new model;

[0034] S(2): Fine-tuning the network parameters of each pre-trained model using the training set of dataset 2;

[0035] S(3): Use the validation set of dataset 2 to determine whether the validation set accuracy meets the requirements. If it does, save the network model; otherwise, return to S(2).

[0036] The beneficial effects of this invention are:

[0037] This invention employs a transfer learning method. When handling astronomical time-series related tasks, the limited verified data from a single telescope and the scarcity of positive samples make it difficult to train deep learning networks. This invention utilizes existing astronomical time-series data acquired by the transit telescope and employs a transfer learning method involving fine-tuning a pre-trained model, using a sufficient number of source domain samples to train the network. After transferring the network structure and parameters to the target domain, the parameters of deeper network layers are fine-tuned, allowing the network to adapt to the data distribution of the target domain samples, thus avoiding overfitting and underfitting caused by insufficient data.

[0038] First, an improved convolutional neural network is pre-trained in the source domain, transferring the network structure and parameters to the target domain. For a given target domain, based on optically variable data with sufficient sample size and label information, a mapping from the source domain to the target domain is established through retraining of the pre-trained model. This avoids underfitting and overfitting caused by small sample sizes and reduces training time, quickly establishing a reliable model for studying data with insufficient samples and labels under different distributions.

[0039] This invention employs a temporal classification method using a one-dimensional convolutional neural network (1DCNN). Compared to the separate processing of feature extraction and classification in machine learning algorithms, the method in this paper does not require manual feature extraction. It can directly apply to the original signal and adaptively extract features beneficial to classification, effectively identifying target attributes. Attached Figure Description

[0040] Figure 1 This is the data preprocessing process of the present invention.

[0041] Figure 2 This is a schematic diagram of the transfer learning algorithm of the present invention.

[0042] Figure 3 This is a flowchart of the transfer learning algorithm of the present invention.

[0043] Figure 4 This is a diagram of the one-dimensional convolutional neural network structure proposed in this invention. Detailed Implementation

[0044] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0045] This invention aims to address the problem of limited datasets in transit-based planetary detection, which hinders model training. It utilizes pre-trained models to accelerate training on future observational data. The invention employs abundant light variation signal data and labels from the source domain to establish a mapping function from samples to labels. By leveraging shared features between the source and target domains, knowledge learned in the source domain can aid in training a classification model in the target domain, enabling knowledge transfer between different domains for exoplanet classification. The process includes: acquiring light variation signals from two telescopes using the transit method for exoplanet detection, using dataset 1 as a pre-training model and dataset 2 as the target model; preprocessing the acquired data; dividing the two datasets into training and testing sets; training a convolutional neural network model using dataset 1 to establish an optimal time-series data classification model; saving the pre-trained model; and adjusting the model using dataset 2 to obtain a model suitable for dataset 2. This method avoids the limitation of models failing to learn with small samples, accelerating data-driven intelligent exoplanet detection.

[0046] This invention includes the following steps:

[0047] The light variation signals of the exoplanets to be classified are input into a pre-trained light variation signal classification model, which outputs the signal classification results. The training process of the light variation signal classification model includes:

[0048] S1: Select two telescopes that observe exoplanets using the transit method. Using a catalog of confirmed exoplanets, acquire light variation signals from each telescope. Select the larger dataset as large sample dataset 1, and the smaller telescope data as small sample dataset 2.

[0049] S2: Segment samples from multiple observation quarters of the same target, and perform missing value processing, outlier processing, motion smoothing, and normalization on the original light variation signal;

[0050] S3: Resample each optical variable signal sample to a fixed length;

[0051] S4: Divide the two datasets into training and validation sets respectively;

[0052] S5: Use dataset 1 to build and train a convolutional neural network model, and save the optimal light variation signal classification model as a pre-trained model;

[0053] S6: Continue training the saved pre-trained model using dataset 2, adjusting it to the optimal light variation signal classification model for the small sample dataset 2.

[0054] The specific process of step S1 is as follows:

[0055] Step S1.1: Download the light variation signals obtained by different transit telescopes through the verified exoplanet catalog. The light variation signal data is the brightness sample value of the observed star over time.

[0056] Step S1.2: For the data collected by the two transit telescopes, use the telescope data with sufficient sample size as a large sample dataset 1 to train the pre-trained model. Use the telescope data with insufficient sample size as a small sample dataset 2.

[0057] The specific process of step S2 is as follows:

[0058] Step S2.1: For the two datasets, concatenate samples from multiple observation quarters of the same target;

[0059] Step S2.2: The missing value processing will delete the luminous flux of all NaN values;

[0060] Step S2.3: The outlier processing will delete outliers that are more than one standard deviation above the median σ1 and less than the median σ2.

[0061] Step S2.4: The motion smoothing process employs Savitzky-Golay weighted motion smoothing filtering, performing k-order polynomial fitting on the data points within the window to remove background noise. This method fits the low-frequency components of the signal, filtering out low-frequency noise while maintaining the signal's shape and width. The formula is as follows:

[0062]

[0063] in, The smoothing coefficient is obtained by fitting a polynomial using the least squares method.

[0064] Step S2.5: The normalization is to eliminate the influence of different feature data units. Before training, the signal sequence is normalized using sample Min-Max normalization, and the formula is as follows:

[0065]

[0066] Where max is the maximum value of the optical variation data and min is the minimum value of the optical variation data.

[0067] The specific process of step S3 is as follows:

[0068] Step S3.1: Use the Discrete Fourier Transform method to resample the spliced ​​signal into n points, with the original sequence interval changed from d. x Become The formula is:

[0069]

[0070] Where x represents the original sequence, X is the result of the discrete Fourier transform of x, n is the number of points in the target sequence, and y is the result of the discrete inverse Fourier transform of X.

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

[0072] Step S4.1: Divide the dataset into N folds for cross-validation, subsequent model evaluation, and model selection.

[0073] The specific process of step S5 is as follows:

[0074] Step S5.1: Use training set 1 to train a 1DCNN network model based on light signal sequences, and select the optimal model structure and model parameters through N-fold cross-validation;

[0075] Step S5.2: Save the 1DCNN model structure and model parameters as a pre-trained model F1 with the suffix ".pth".

[0076] The specific process of step S6 is as follows:

[0077] Read the pre-trained model file F1 to obtain the model structure and its parameters, and use the pre-trained model parameters as the initial values ​​for the new model parameters. Fine-tune the network structure and parameters of the pre-trained model using the training set of Dataset 2, and test the new model using the validation set of Dataset 2. Finally, through N-fold cross-validation, select the best model structure and parameters suitable for this small sample size.

[0078] The following is an embodiment with reference to the accompanying drawings:

[0079] Part 1: Acquisition and processing flow of astronomical time series signal data.

[0080] In the search for exoplanets, the Kepler Space Telescope has become a major data provider. The observational data processed by the Kepler Pipeline laid the foundation for the application of deep learning in exoplanet hunting.

[0081] This invention can select data from transit surveys with large existing datasets, such as Kepler and TESS, as a large sample dataset 1. Other telescope data with smaller datasets can be selected as a small sample dataset 2, and then divided into training and validation sets respectively.

[0082] like Figure 1 After acquiring the light variation signal, samples from multiple observation seasons of the same target are spliced ​​together;

[0083] Perform missing value processing and remove luminous flux with all NaN values.

[0084] Outlier removal involves deleting outliers that are more than one standard deviation above the median σ1 and less than the median σ2.

[0085] Moving-mode smoothing is performed using a Savitzky-Golay weighted moving-mode smoothing filter. A k-th order polynomial is fitted to the data points within the window to remove background noise. This method fits the low-frequency components of the signal, filtering out low-frequency noise while preserving the signal's shape and width. The formula is as follows:

[0086]

[0087] in, The smoothing coefficient is obtained by fitting a polynomial using the least squares method.

[0088] Normalization is used to eliminate the influence of different feature data units. Before training, the signal sequence is normalized using Min-Max normalization, and the formula is as follows:

[0089]

[0090] Where max is the maximum value of the optical variation data and min is the minimum value of the optical variation data.

[0091] The discrete Fourier method is used to resample the spliced ​​signal into n points, with the original sequence interval changed from d. x Become The formula is:

[0092]

[0093] Where x represents the original sequence, X is the result of the discrete Fourier transform of x, n is the number of points in the target sequence, and y is the result of the discrete inverse Fourier transform of X.

[0094] For the two datasets, each dataset is divided into N folds for cross-validation, subsequent model evaluation, and model selection.

[0095] Part Two: Obtaining the Source Domain Pre-trained Model.

[0096] Training set 1 was used to train a 1DCNN network model based on optically variable signal sequences, and the optimal model structure and parameters were selected as the pre-trained model through N-fold cross-validation. Figure 2 .

[0097] like Figure 3 and Figure 4 During the model training process in the source domain, the parameters of the neural network are first initialized. After the network forward propagates through alternating convolutional and pooling layers, fully connected layers, and softmax layers, the network error is calculated, and it is determined whether the error meets the requirements. If not, the network parameters are updated by backpropagating the error using an optimization algorithm until the network error or the number of iterations reaches the set requirements.

[0098] Determine whether the accuracy of the final validation set meets the requirements. If it does, save the network model. If not, reconfigure the network's hyperparameters and repeat the above steps.

[0099] Save the optimal 1DCNN model structure and model parameters as a pre-trained F1 model with the suffix ".pth".

[0100] Part 3: Migrating the source domain model to the target domain.

[0101] like Figure 3 In the target domain, the pre-trained model file F1 is read to obtain the model structure and its parameters. The pre-trained model parameters are used as the initial values ​​for the new model parameters. The network structure and parameters of each pre-trained model are fine-tuned using the training set of Dataset 2, and the new model is tested using the validation set of Dataset 2. Finally, through N-fold cross-validation, the optimal model structure and parameters suitable for this small sample size are selected.

[0102] It should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0103] The descriptions of the embodiments herein are merely illustrative of the methods of the present invention and represent only a portion, not all, of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention, and the content of this specification should not be construed as a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for classifying exoplanet light variation signals based on transfer learning, characterized in that: The light variation signals of the exoplanets to be classified are input into a pre-trained light variation signal classification model, and the signal classification results are output. The optical signal classification model employs a one-dimensional convolutional neural network, including alternating convolutional and pooling layers, fully connected layers, and softmax layers. The training method includes: S1: Select two telescopes that use the transit method to observe exoplanets. Obtain light variation signals from different telescopes through a catalog of verified exoplanets. Select the telescope data with larger data volume as large sample dataset 1 and the telescope data with smaller data volume as small sample dataset 2. S2: In two datasets, samples from multiple observation quarters of the same target are spliced ​​together. The spliced ​​light variation signal from the two datasets is then processed for missing values, outliers, moving smoothing, and normalization. S3: Resample each optical variable signal sample to a fixed length; S4: Divide both dataset 1 and dataset 2 into training set and validation set; S5: Use dataset 1 to train the constructed one-dimensional convolutional neural network and save the optimal light-varying signal classification model as the source domain pre-training model; S6: Continue training the pre-trained model using dataset 2, transfer the source domain model to the target domain, and adjust to obtain the optimal light variation signal classification model for the small sample dataset 2.

2. The method for classifying exoplanet light variation signals based on transfer learning according to claim 1, characterized in that: The outlier removal process involves deleting outliers that are more than one standard deviation above the median σ1 and less than the median σ2.

3. The method for classifying exoplanet light variation signals based on transfer learning according to claim 1, characterized in that: The motion smoothing process includes: A weighted moving average smoothing filter, Savitzky-Golay, is used to perform k-order polynomial fitting on the data points within the window to remove background noise. in, The smoothing coefficient is obtained by fitting a polynomial using the least squares method.

4. The method for classifying exoplanet light variation signals based on transfer learning according to claim 1, characterized in that: The normalization process includes: Perform Min-Max normalization on the samples: Where max is the maximum value of the optical variation data and min is the minimum value of the optical variation data.

5. The method for classifying exoplanet light variation signals based on transfer learning according to claim 1, characterized in that: S3 describes resampling each optically variable signal sample to a fixed length, which includes: The discrete Fourier method is used to resample the spliced ​​signal into n points, with the original sequence interval changed from d. x Become Specifically: X = dft(x) y = idft(X, n) Where x represents the original sequence, X is the result of the discrete Fourier transform of x, n is the number of points in the target sequence, and y is the result of the discrete inverse Fourier transform of X.

6. The method for classifying exoplanet light variation signals based on transfer learning according to claim 1, characterized in that: The training of the constructed one-dimensional convolutional neural network using dataset 1 includes: S1: Initialize the parameters of the one-dimensional convolutional neural network; S2: After the network forward propagation passes through alternating convolutional and pooling layers, fully connected layers, and softmax layers, the network error is calculated, and it is determined whether the error meets the requirements. If not, the network is updated by adjusting the network combination and hyperparameters until the network error or the number of iterations reaches the set requirements. S3: Determine if the accuracy of the validation set meets the requirements. If it does, save the network model; otherwise, reconfigure the network structure and hyperparameters, and return to S2.

7. The method for classifying exoplanet light variation signals based on transfer learning according to claim 1, characterized in that: The step of continuing to train the pre-trained model using dataset 2 and transferring the source domain model to the target domain includes: S(1): Transfer the network structure and parameters of the pre-trained model to the new model, and use the parameters of the pre-trained model as the initial values ​​of the parameters of the new model; S(2): Fine-tuning the network parameters of each pre-trained model using the training set of dataset 2; S(3): Use the validation set of dataset 2 to determine whether the validation set accuracy meets the requirements. If it does, save the network model; otherwise, return to S(2).

Images