A method and system for detecting Jupiter magnetic reconnection events based on deep learning algorithm

Abstract

The present invention belongs to the application research of machine learning technology in the field of space physics, and specifically relates to a method and system for detecting Jupiter magnetic reconnection events based on a deep learning algorithm, the method comprising: obtaining Jupiter magnetic field observation data based on time periods and performing preprocessing; inputting the preprocessed data into a pre-established and trained magnetic reconnection event detection model, and outputting a detection result of whether a specific time period contains Jupiter magnetic reconnection events; the magnetic reconnection event detection model uses a 1DResNet residual network to mine magnetic field observation data, realize a semantic reasoning process from bottom to top, and identify the regional blocks containing Jupiter magnetic reconnection in each regional block in the time series magnetic field data, thereby realizing the detection of magnetic reconnection events. The method of the present invention uses an end-to-end approach and is suitable for automatic analysis and processing of a large amount of Jupiter magnetic field data.

Description

Technical Field

[0001] The present invention relates to the application research of machine learning technology in the field of space physics, and specifically to a method and system for detecting Jupiter magnetic reconnection events based on a deep learning algorithm. Background Art

[0002] Magnetic reconnection, a physical process in which magnetic field energy is rapidly converted into plasma kinetic energy, occurs frequently in planetary space and is generally believed to be closely related to particle acceleration and planetary magnetic storms. Satellite observations of Earth's magnetosphere have yielded numerous magnetic reconnection events. With the implementation of various deep-space planetary exploration satellite programs, an increasing amount of planetary magnetosphere observation data is being accumulated, and scientists have discovered magnetic reconnection events similar to those observed in Earth's magnetosphere on various planets.

[0003] Magnetic reconnection events in the magnetosphere of different planets share similar physical characteristics. However, due to differences in size and magnetic field strength, each planet also exhibits unique characteristics. The primary purpose of this invention is to detect and discover magnetic reconnection events on Jupiter. Currently, Jupiter's magnetosphere observation data primarily comes from the Galileo and Juno satellites. Juno has been observing the Jupiter magnetosphere since arriving at Jupiter in 2016. Its orbital period is approximately 53 Earth days, and its aphelion is approximately 113 Jupiter radii. Vogt et al. (2010) and Vogt et al. (2019) list over 200 manually identified magnetic reconnection events observed by the Galileo and Juno satellites. Through comparison, they identify the unique characteristics of their Jupiter observations.

[0004] Traditionally, magnetic reconnection events have been identified visually, using set criteria to identify them within vast amounts of satellite observational time series data. This process consumes considerable time and effort, is subject to considerable subjectivity, and introduces errors. Summary of the Invention

[0005] In view of the problems of low manual identification efficiency and subjectivity in the existing methods for detecting Jupiter's magnetic reconnection events, as well as difficulty in processing massive amounts of Jupiter's magnetic field data, the purpose of the present invention is to overcome the above-mentioned shortcomings of the existing technology and propose a method and system for detecting Jupiter's magnetic reconnection events based on a deep learning algorithm.

[0006] In view of this, the present invention proposes a method for detecting Jupiter magnetic reconnection events based on a deep learning algorithm, comprising:

[0007] Obtain Jupiter's magnetic field observation data based on time periods and perform preprocessing;

[0008] The pre-processed data is input into a pre-established and trained magnetic reconnection event detection model, which outputs the detection result of whether a specific time period contains a Jupiter magnetic reconnection event;

[0009] The magnetic reconnection event detection model adopts the 1DResNet residual network to mine magnetic field observation data, realize the semantic reasoning process from bottom to top, and identify the regional blocks containing Jupiter's magnetic reconnection in each regional block in the time series magnetic field data, thereby realizing the detection of magnetic reconnection events.

[0010] Preferably, the time resolution of the magnetic field observation data is 1 second.

[0011] Preferably, the method includes a step of training a magnetic reconnection event detection model, comprising:

[0012] Collect data, perform preprocessing, and build a training set;

[0013] The data of the training set are input into the 1DResNet residual network in sequence for supervised learning. The loss function uses cross entropy until the training requirements are met, and a trained magnetic reconnection event detection model is obtained.

[0014] Preferably, the data collection includes: selecting original observation data from the magnetic field observation instrument FGM of the Juno satellite.

[0015] Preferably, the preprocessing includes: data outlier processing, missing value filling, magnetic field spherical coordinate conversion and background average magnetic field removal.

[0016] Preferably, a linear interpolation method is used to process data outliers and fill in missing values.

[0017] Preferably, the magnetic field spherical coordinate transformation includes:

[0018] The magnetic field observation data is converted from the magnetic field observation components in the three directions of x, y, and z in the Cartesian coordinate system at the center of Jupiter 、 and , converted into three magnetic field components in spherical coordinates 、 and , satisfying the following formula:

[0019]

[0020]

[0021]

[0022] in, is the pitch angle, is the azimuth, r is the distance from the center of Jupiter, and x, y, and z represent the position components of the observation point, respectively, satisfying the following formula:

[0023]

[0024] When x>0, ;when hour, .

[0025] Preferably, the background average magnetic field removal includes:

[0026] Before calculation, the average magnetic field of the set hour is taken as the background magnetic field, and the background magnetic field is subtracted from the magnetic field value of the observation time period.

[0027] Preferably, establishing a training set includes:

[0028] The pre-processed data are annotated with Jupiter's magnetic reconnection events, including the start and end time, start and end position, and duration of each magnetic reconnection event;

[0029] The labeled data with Jupiter magnetic reconnection events are sliced ​​according to the time slice pairs to obtain the magnetic reconnection event data segments and corresponding labels to form positive samples;

[0030] Corresponding time slices are intercepted, and the data and corresponding labels of Jupiter's magnetic reconnection events are not included to form negative samples; thus, a training set is obtained.

[0031] On the other hand, the present invention provides a Jupiter magnetic reconnection event detection system based on a deep learning algorithm, comprising a preprocessing module and a detection output module, wherein:

[0032] The preprocessing module is used to obtain Jupiter's magnetic field observation data based on time periods and perform preprocessing;

[0033] The detection output module is used to input the pre-processed data into a pre-established and trained magnetic reconnection event detection model, and output a detection result of whether a specific time period contains a Jupiter magnetic reconnection event;

[0034] The magnetic reconnection event detection model adopts the 1DResNet residual network to mine magnetic field observation data, realize the semantic reasoning process from bottom to top, and identify the regional blocks containing Jupiter's magnetic reconnection in each regional block in the time series magnetic field data, thereby realizing the detection of magnetic reconnection events.

[0035] Compared with the prior art, the advantages of the present invention are:

[0036] Based on the typical characteristics of Jupiter's magnetic reconnection events, this paper proposes a Jupiter magnetic reconnection event recognition model based on a supervised deep neural network model (DResNet). This model implements a bottom-up reasoning process and ultimately extracts the observation time period containing Jupiter magnetic reconnection events. This proposed technical solution utilizes an end-to-end approach. After the model is trained, no parameters need to be adjusted. Simply inputting the data for the corresponding time period directly outputs the corresponding identification results. This makes it suitable for the automated analysis and processing of large amounts of Jupiter magnetic field data. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a schematic diagram of the spherical coordinate transformation method;

[0038] Figure 2 This is a flow chart of the method for detecting Jupiter magnetic reconnection events of the present invention;

[0039] Figure 3 This is an example of a major magnetic reconnection event of Jupiter observed by the JUNO probe;

[0040] Figure 4 is a positive label instance (reconnection event occurs);

[0041] Figure 5 It is the local connection structure of the residual network;

[0042] Figure 6 This is an example of magnetic reconnection detection at Jupiter. DETAILED DESCRIPTION

[0043] With the tremendous success and widespread application of deep learning neural networks in extracting features from various image and time series data, researchers in the fields of space weather and planetary science have also used machine learning techniques to analyze various image and time series data. The search and identification of magnetic reconnection events is a typical task in identifying characteristic events from time series data. Therefore, based on this need, this paper primarily utilizes algorithms for processing time series data features in deep learning. Based on magnetic field observations from the Juno satellite and the manually selected magnetic reconnection events described in the paper by Vogt et al. (2019), we develop a set of recognition algorithms for detecting Jupiter magnetic reconnection events. This improves the automated batch processing capabilities of Jupiter magnetic field observation data and facilitates statistical analysis of Jupiter magnetic reconnection event data. This paper constructs a Jupiter magnetic reconnection event detection method based on deep learning technology, enabling the rapid and accurate automatic extraction of Jupiter magnetic reconnection event regions from long-term observational data, meeting the statistical analysis needs of Jupiter magnetic field and magnetic reconnection researchers for long-term, large-sample statistical analysis.

[0044] like Figure 2 As shown, this technical solution mainly includes the following steps:

[0045] 1. Data collection: Select the original observation data of the fluxgate magnetometer (FGM), the magnetic field observation instrument of the Juno satellite.

[0046] 2. Data preprocessing: including data outlier processing, missing value filling, magnetic field spherical coordinate transformation, background average magnetic field removal and other data preprocessing;

[0047] 3. Dataset creation: Manually screen samples containing magnetic reconnection events and annotate their start and end times to construct the annotated dataset required for training the magnetic reconnection event recognition model;

[0048] 4. Model selection and testing: Verify the applicability of the model to Juno magnetic field data, and select a network model structure with better performance based on the characteristics of magnetic reconnection events and FGM magnetic field observation data.

[0049] 5. Model evaluation: Use the labeled dataset sample data as the input of the selected magnetic reconnection event recognition model, adjust the parameters to train the recognition model, obtain the model recognition results and evaluate them.

[0050] Below we further explain the specific steps:

[0051] Step 1 involves data collection. First, based on the literature and the FGM data from the JUNO satellite magnetic field instrument, we screened manually annotated magnetic reconnection events from 2016 to 2019 and the corresponding magnetic field data files. Magnetic field data is typically stored as a single file per day, with a time resolution of 1 second. The file header primarily records auxiliary observation parameters and coordinate system characteristics. The physical meaning and storage format of each observed parameter are also detailed.

[0052] Step 2 is data preprocessing, including data outlier processing, missing value filling, magnetic field spherical coordinate conversion, background average magnetic field removal, etc. Since the original data may have abnormal observation points at some times or be missing in some time periods (short time) during the operation of the observation instrument, it is necessary to process the outliers and missing values ​​of the data through linear interpolation and other methods. The rationality of the magnetic field data is guaranteed by time resampling. Since the actual judgment of Jupiter's magnetic reconnection events mainly relies on the tangential component of the pitch angle of the magnetic field, it is necessary to convert the magnetic field observation data from the components in the x, y, and z directions of the Cartesian coordinate system at the center of Jupiter. 、 and , converted into three magnetic field components in spherical coordinates 、 and , are unit vectors in the corresponding directions respectively.

[0053] like Figure 1 The following is a schematic diagram of the spherical coordinate conversion method. Subsequent calculations are based on the three components of the spherical coordinate system to calculate the corresponding features. The conversion method and formula are as follows:

[0054]

[0055]

[0056]

[0057] in, is the pitch angle, is the azimuth, satisfying the following formulas:

[0058]

[0059]

[0060] When x>0, ;when hour, .

[0061] At the same time, magnetic reconnection events are mainly determined by observing the disturbance of the magnetic field. Therefore, it is often necessary to calculate the disturbance equivalent to the background magnetic field. Here, the magnetic field calm value 10 hours before the calm period is used as the standard for the background magnetic field.

[0062] Step 3 involved dataset creation. By collecting annotations of Jupiter's magnetic reconnection events from publicly available papers and manually correcting them, we initially selected 233 Jupiter magnetic reconnection events from 2016 to 2019 and annotated each event with the features required for this study. This primarily recorded the start and end times, locations, and duration of each magnetic reconnection event. Based on the corresponding magnetic reconnection event times, we sliced ​​the corresponding time segments using magnetic field data to obtain the corresponding magnetic reconnection event data segments and corresponding labels (positive samples). To meet the requirements of subsequent detection methods, we also required a certain number of negative samples. To this end, we intercepted data of equal durations that did not contain magnetic reconnection events and their corresponding labels to create negative samples.

[0063] Step 4 involves model selection and testing. Considering the characteristics of magnetic reconnection events and magnetic field time-series observations, the 1DResNet residual network, known for its excellent performance in multi-scale feature fusion, was selected as the network model for detecting Jupiter's magnetic reconnection events. This model improves the performance of deep neural networks by introducing improvements such as span fusion of features across each layer. Furthermore, this network model was experimentally compared with other commonly used one-dimensional feature network models to verify the advantages of the 1DResNet model in detecting Jupiter's magnetic reconnection events.

[0064] Step 5 involves model evaluation. The labeled dataset samples are used as input for the selected Jupiter magnetic reconnection detection model. The model is trained by adjusting parameters, obtaining model extraction results and conducting an evaluation. The confusion matrix, a commonly used evaluation metric in the classification field, is used to measure model performance. This task primarily involves identifying whether a piece of time series data represents a magnetic reconnection event, a typical binary classification problem. In binary classification problems, the confusion matrix has four basic components: True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN).

[0065] True Positive (TP) is when the model predicts the positive class and the prediction is correct, we call it True Positive.

[0066] True Negative (TN) is when the model predicts a negative class, and the prediction is correct, we call it True Negative.

[0067] False Positive (FP) is when the model predicts a positive class, but the prediction is wrong, which we call a False Positive.

[0068] False Negative (FN) is when the model predicts a negative class, but the prediction is wrong, which we call a False Negative.

[0069] Precision is the ratio of the number of samples that the model predicts to be positive and are actually positive to the number of samples that the model predicts to be positive. .

[0070] Recall is the ratio of the number of samples predicted by the model to be positive and actually positive to the number of actual positive samples. Since magnetic reconnection events are rare events in the entire Jupiter magnetosphere, precision and recall are relatively important indicators in this mission.

[0071] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0072] Example 1

[0073] Example 1 of the present invention proposes a method for detecting Jupiter magnetic reconnection events based on a deep learning algorithm, and its research and development technical route is as follows:

[0074] (1) First, based on the literature and the FGM data of the JUNO satellite magnetic field instrument to be applied, manually annotated magnetic reconnection events from 2016 to 2019 and the corresponding magnetic field data files were screened. In order to construct an intelligent recognition method, according to the physical laws of the magnetic reconnection phenomenon and the characteristics of the data, a certain number of preprocessing steps are required for the data, mainly including missing data point completion, spherical coordinate vector conversion, and background magnetic field subtraction.

[0075] like Figure 3 Shown is an example of the main magnetic reconnection events of Jupiter observed by the JUNO probe.

[0076] The table above primarily records the start and end time, start and end location, and duration of each magnetic reconnection event. Based on the corresponding magnetic reconnection event time, we sliced ​​the corresponding time segments using magnetic field data to obtain the corresponding magnetic reconnection event data segments and corresponding labels (positive samples). To meet the needs of subsequent detection methods, we also required a certain amount of negative samples. To this end, we intercepted data of similar durations that did not contain magnetic reconnection events and their corresponding labels to create negative samples. In this study, due to the limited availability of Jupiter data and magnetic reconnection events, to ensure the reliability and stability of the algorithm, we did not split the dataset and test sets separately. Instead, we used a three-fold cross-validation method. This method divides the entire sample set into three equal parts. For each training and testing run, two equal parts are used as the training set, and the remaining one equal part is used as the validation set. This method generates an accuracy rate for each training run, and the three-fold validation is repeated three times, with each equal part used as the validation set. This results in an average accuracy rate that closely matches the actual performance of the model algorithm.

[0077] like Figure 4 The following are typical examples of positive and negative samples.

[0078] (2) After completing the data cleaning process, the main classifier algorithm is constructed. The main applications are the traditional empirical mode decomposition analysis method and the deep learning network analysis method. The corresponding classifiers and judgment criteria are constructed respectively to determine whether the magnetic reconnection event occurs within the target time.

[0079] Empirical mode decomposition (EMD) is based on the concept of instantaneous frequency intrinsic mode functions (IMFs). It can decompose complex signals into several intrinsic mode functions (IMFs), each representing a local feature of the signal. Signal decomposition is performed based on the time-scale characteristics of the data. The basic idea is to transform a wave with irregular frequency into multiple single-frequency waves and residual waves. In this detection algorithm, the training set sample labels are first taken, and the magnetic field data is deducted from the average background over the past 10 hours. After obtaining the changing waveform, empirical mode decomposition is performed to obtain five single-frequency waves and residual waves. Then, in the corresponding detection time region, the corresponding three magnetic field components are subjected to empirical mode decomposition. The labels are then compared to calculate the similarity distance. Finally, based on a set similarity distance threshold, the region is determined to determine whether it contains a magnetic reconnection event. The corresponding recall and precision rates are then calculated.

[0080] 1DResnet is a 1D application based on the classic deep learning algorithm Resnet50. Similarly, we first need to use the public pre-trained model based on 1DResnet in our task. This task uses a 2-classification task, that is, to determine whether the corresponding time segment contains a magnetic reconnection event, which is a positive / negative sample judgment problem. Unlike traditional methods, the method based on deep learning 1DResnet also needs to train negative samples, and Loss uses cross entropy, which is suitable for 2-classification tasks. The ResNet network refers to the VGG19 network, and is modified on its basis, and residual units are added through a short-circuit mechanism, such as Figure 5 As shown in the figure, x is the input variable of each layer of the residual network, weight layer is the weight parameter of each layer of the model, F(x) is the result after the weight calculation of this layer, and is finally added to the original variable (identity) x to form the output of this layer. Relu is the activation function, which is the commonly used linear rectification function.

[0081] In order to verify the reliability of the classifier construction algorithm, the experimental results of the classification need to be cross-validated, that is, to eliminate the bias of manually dividing the training set and the validation set. After obtaining reliable results, the constructed intelligent recognition model is used to perform forward inference of the model, screen out more magnetic reconnection events, and manually verify the screened events. Screening out Jupiter's magnetic reconnection events through intelligent recognition algorithms greatly saves the energy consumed by scientists in manual judgment, while making up for the errors caused by human subjective factors, and providing support for subsequent research on Jupiter's magnetic reconnection events. After experimental testing, the accuracy of Jupiter's magnetic reconnection event detection obtained by the traditional method of empirical mode decomposition was 78%, the recall rate was 80.8%, and the F1 value reached 79.5%. The accuracy of Jupiter's magnetic reconnection event detection obtained by the deep learning 1DResnet method was 82.1%, the recall rate was 77%, and the F1 value reached 79.4%. The following are typical magnetic reconnection event detection results. Figure 6 As shown. This is the event detection on the 195th day of 2016. The time span in the figure is 1800-2300UT. The four columns in the figure respectively draw the three components B of the magnetic field vector observed by the FGM instrument in the spherical coordinate system after coordinate transformation. r ,B θ ,B φ and the total magnetic field size B t The corresponding detected magnetic reconnection region is shown in pink shading. The magnetic reconnection region occurred roughly from 1952 to 2049 UT. As can be seen from the figure, the changing characteristics of the magnetic field components within the region of magnetic reconnection are similar to those of the positive samples in the figure above, with a corresponding bipolar structure, which is consistent with theoretical experience. This shows that the corresponding detection algorithm has basically achieved the expected level of performance in detecting typical magnetic reconnection events.

[0082] Example 2

[0083] Embodiment 2 of the present invention provides a Jupiter magnetic reconnection event detection system based on a deep learning algorithm, which is implemented based on the method of embodiment 1. The system includes a preprocessing module and a detection output module, wherein:

[0084] The preprocessing module is used to obtain Jupiter's magnetic field observation data based on time periods and perform preprocessing;

[0085] The detection output module is used to input the pre-processed data into a pre-established and trained magnetic reconnection event detection model, and output a detection result of whether a specific time period contains a Jupiter magnetic reconnection event;

[0086] The magnetic reconnection event detection model adopts the 1DResNet residual network to mine magnetic field observation data, realize the semantic reasoning process from bottom to top, and identify the regional blocks containing Jupiter's magnetic reconnection in each regional block in the time series magnetic field data, thereby realizing the detection of magnetic reconnection events.

[0087] Innovation:

[0088] This technical solution introduces deep learning technology for the first time to conduct information mining on massive Jupiter magnetic field data, realize the semantic reasoning process from bottom to top, identify each regional block in the time series magnetic field data that contains Jupiter magnetic reconnection and extract it.

[0089] The model is highly robust. After training, no parameters need to be adjusted. Simply input the corresponding time period data into the model, and it will directly output whether the corresponding time period contains Jupiter's magnetic reconnection event. Therefore, it is suitable for automatic analysis and processing of large amounts of time series data.

[0090] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.

Claims

1. A method for detecting Jupiter magnetic reconnection events based on a deep learning algorithm, comprising: Obtain Jupiter's magnetic field observation data based on time periods and perform preprocessing; The pre-processed time-series magnetic field observation data is input into a pre-established and trained magnetic reconnection event detection model, and the detection result of whether the time period contains the Jupiter magnetic reconnection event is output; The magnetic reconnection event detection model adopts 1DResNet residual network to mine the time series magnetic field observation data, realize the semantic reasoning process from bottom to top, and identify the regional blocks containing Jupiter magnetic reconnection in each regional block in the time series magnetic field observation data, so as to realize the detection of magnetic reconnection events; The method includes a step of training a magnetic reconnection event detection model, including: Collect data, perform preprocessing, and build a training set; The data of the training set are sequentially input into the 1DResNet residual network for supervised learning, and the loss function adopts cross entropy until the training requirements are met to obtain a trained magnetic reconnection event detection model; The data collection includes: selecting the original observation data of the magnetic field observation instrument FGM of the Juno satellite; The preprocessing includes: data outlier processing, missing value filling, magnetic field spherical coordinate transformation and background average magnetic field removal; The background average magnetic field removal includes: The average magnetic field in the set hour before calculation is taken as the background magnetic field, and the background magnetic field is subtracted from the magnetic field value during the observation period.

2. The Jupiter magnetic reconnection event detection method based on deep learning algorithm according to claim 1 is characterized in that: The time resolution of the magnetic field observation data is 1 second.

3. The Jupiter magnetic reconnection event detection method based on deep learning algorithm according to claim 1, characterized in that: Linear interpolation methods are used to process data outliers and fill in missing values.

4. The Jupiter magnetic reconnection event detection method based on deep learning algorithm according to claim 1, characterized in that: The magnetic field spherical coordinate transformation includes: The magnetic field observation data is converted from the components of the x, y, and z directions of the Cartesian coordinate system centered on Jupiter , and , converted into three magnetic field components in spherical coordinates , and , satisfying the following formula: ； ； ； in, is the pitch angle, is the azimuth, r is the distance from the center of Jupiter, x, y, z represent the position components of the observation point, respectively, satisfying the following formula: ； When x>0, ;when hour, .

5. The Jupiter magnetic reconnection event detection method based on deep learning algorithm according to claim 1, characterized in that: The establishing of the training set comprises: The pre-processed data are annotated with Jupiter's magnetic reconnection events, including the start and end time, start and end position, and duration of each magnetic reconnection event; Slice the annotated data with Jupiter magnetic reconnection events according to the time slice pairs to obtain the magnetic reconnection event data fragments and corresponding labels to form positive samples; Corresponding time slices of equal length are intercepted, and data and corresponding labels without Jupiter magnetic reconnection events are taken to form negative samples; thus, a training set is obtained.

6. A system based on the Jupiter magnetic reconnection event detection method based on deep learning algorithm as claimed in claim 1, characterized in that: It includes a preprocessing module and a detection output module, wherein: The preprocessing module is used to obtain the magnetic field observation data of Jupiter based on the time period and perform preprocessing; The detection output module is used to input the pre-processed data into a pre-established and trained magnetic reconnection event detection model, and output a detection result of whether a specific time period contains a Jupiter magnetic reconnection event; The magnetic reconnection event detection model adopts the 1DResNet residual network to mine the magnetic field observation data, realize the semantic reasoning process from the bottom layer to the high layer, and identify the regional blocks containing Jupiter's magnetic reconnection in each regional block in the time series magnetic field data, thereby realizing the detection of magnetic reconnection events.

Images