Learning device, learning method, learning program, information processing system, information processing method, program, and recording medium

Abstract

To provide a learning device capable of constructing a machine learning model capable of accurately predicting electrification of an electrification prediction target object.SOLUTION: The learning device includes a data input unit and a control unit. The data input unit receives a learning data set including solar wind data and charging information of the artificial satellite to be observed, which are associated with each other. The control unit constructs a machine learning model that outputs a possibility of electrification of the electrification prediction target object from the input solar wind data by performing machine learning using the learning data set. The electrification information includes information of a surface electrification label indicating whether or not the surface state of the observation artificial satellite is electrified, which is given based on data of an energy spectrogram related to an ion flux and / or an electron flux observed by the observation artificial satellite.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a learning device, a learning method, a learning program, an information processing system, an information processing method, a program, and a recording medium. [Background technology]

[0002] In satellite missions that involve docking operations between a new satellite and a satellite already in orbit for purposes such as space debris removal and in-orbit servicing, there is a risk that discharges caused by potential differences between the satellites could damage the satellite's systems and equipment.

[0003] Countermeasures against the potential difference include providing anti-charging measures to space debris capture systems, mitigating charging using plasma generators such as electric thrusters (see, for example, Patent Document 1 and Non-Patent Document 1), on-board measurement using charge measurement devices, etc. Patent Document 2 also describes a machine learning method for building a space weather forecast system. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] U.S. Patent No. 9,119,277 [Patent Document 2] Japanese Patent Application Laid-Open No. 2016-31282 [Non-patent literature]

[0005] [Non-Patent Document 1] NASA / TM-2014-216579 Operational Status of the International Space Station Plasma Contactor Hollow Cathode Assemblies From July 2-11 to May 2013 Summary of the Invention [Problem to be solved by the invention]

[0006] Measures such as adding anti-charging measures to the capture system, mitigating charging using plasma generators such as electric thrusters, and on-board measurement using charge measurement devices all place a burden on the satellite system.

[0007] An object of the present invention is to provide a learning device, a learning method, a learning program, an information processing system, an information processing method, a program, and a recording medium that are capable of predicting the charging of an artificial satellite. [Means for solving the problem]

[0008] A learning device according to an embodiment of the present invention includes a data input unit and a control unit. The data input unit receives input of a learning data set of solar wind data and electrostatic charge information of an observation satellite, which are associated with each other. The control unit performs machine learning using the learning data set, and constructs a machine learning model that outputs the possibility of charging of an object to be predicted to be charged, based on the input solar wind data. The charge information includes information on a surface charge label that indicates whether the surface state of the observation satellite is charged or not, and is assigned based on energy spectrogram data related to ion flux and / or electron flux observed by the observation satellite.

[0009] With this configuration, it is possible to construct a machine learning model that can accurately predict the possibility of charging an object to be predicted to be charged.

[0010] A learning method according to another aspect of the present invention is a learning method for constructing a machine learning model that outputs the possibility of charging of an object for which charging prediction is to be performed, based on input solar wind data, by performing machine learning using a learning data set of solar wind data and charging information from an observation satellite that are associated with each other. The charge information includes information on a surface charge label that indicates whether the surface state of the observation satellite is charged or not, and is assigned based on energy spectrogram data related to ion flux and / or electron flux observed by the observation satellite.

[0011] A learning program according to yet another aspect of the present invention includes: A step of constructing a machine learning model that outputs the possibility of charging of an object to be predicted from input solar wind data by performing machine learning using a learning dataset of solar wind data and charging information from observation satellites that are associated with each other. It is a learning program that allows you to: The charge information includes information on a surface charge label that indicates whether the surface state of the observation satellite is charged or not, and is assigned based on energy spectrogram data related to ion flux and / or electron flux observed by the observation satellite.

[0012] An information processing system according to yet another embodiment of the present invention includes a control unit. The control unit acquires solar wind data and calculates the possibility of charging of the object to be predicted to be charged from the solar wind data using the trained machine learning model constructed by the learning device described above.

[0013] An information processing method according to yet another aspect of the present invention has an information processing device execute the steps of acquiring solar wind data and calculating the possibility of charging an object to be predicted to be charged from the solar wind data using the trained machine learning model constructed by the learning device described above.

[0014] A program according to yet another aspect of the present invention causes an information processing device to execute the steps of acquiring solar wind data and predicting the charge of an object to be predicted to be charged from the solar wind data using the trained machine learning model constructed by the learning device described above.

[0015] A recording medium according to yet another embodiment of the present invention is a non-transitory computer-readable recording medium storing the above program. [Effects of the Invention]

[0016] According to a learning device according to an aspect of the present invention, it is possible to construct a machine learning model that can accurately predict the charge of an object to be predicted for charge prediction. [Brief explanation of the drawings]

[0017] [Figure 1] 1 is a functional block diagram of a learning system including a learning device according to an embodiment of the present invention. [Figure 2] FIG. 2 is a hardware configuration diagram of the learning device. [Figure 3] This is an example of an energy spectrogram of ions and electrons observed by the DMSP satellite, used when creating training data. [Figure 4] 1 is a diagram showing a schematic diagram of the positional relationship between the sun, a solar wind observation satellite, and the earth. [Figure 5] The distribution of the AE Index (Auroral Electrojet Index) when charged and when not charged is shown. [Figure 6] (A) shows the distribution of charged and uncharged states at geomagnetic latitudes in the training data set, and (B) shows the distribution of charged and uncharged states at local time in the training data set. [Figure 7] 10 is a flow chart showing a method for creating a learning dataset using the learning device. [Figure 8] 1 is a flow chart of a learning method for constructing a machine learning model using the learning device. [Figure 9]This shows the training dataset (true label) and the machine learning model output (logit) for the energy-time spectrogram of ion differential energy flux and electron differential energy flux obtained by the DMSP-F18 satellite on January 3, 2022. [Figure 10] The confusion matrix for validation data used to verify the accuracy of machine learning classification results when the threshold is set to 0.5 is shown. [Figure 11] FIG. 10 is a functional block diagram of an information processing system including an information processing device that performs information processing using a trained machine learning model trained by the learning device. [Figure 12] 10 is a flow chart of an information processing method using a trained machine learning model by the information processing device. [Figure 13] FIG. 10 is an example of a display image generated by the information processing device, which visualizes the possibility of electrification in the auroral zone. [Figure 14] This shows a histogram of the machine learning model output for each label on the validation data. [Figure 15] This shows the confusion matrix for validation data used to verify the accuracy of machine learning classification results when the threshold is set to 0.5, and when the likelihood output from the model is limited to samples with a likelihood of 0.2 or less or 0.8 or more. [Figure 16] 10 is a graph illustrating the performance of a trained machine learning model according to another embodiment. [Figure 17] 1 is a graph showing the distribution of AE INDEX for each surface charging potential. [Figure 18] FIG. 10 is a diagram showing the importance of features used in training a machine learning model, with gain as an index. [Figure 19] 10 is a flow chart showing a training dataset creation method according to yet another embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0018] <Summary of the Invention>

[0019] Hereinafter, embodiments of the present invention will be described with reference to the drawings. The following first to third embodiments relate to a learning device that constructs a machine learning model that predicts the charge of an object to be predicted to be charged, and an information processing system that uses the trained machine learning model constructed by the learning device. In the following embodiments, the description will be given taking as an example the charge prediction of an object to be predicted to be charged in the aurora zone, but the present invention is not limited to the aurora zone, and can also be applied to the charge prediction of an object to be predicted to be charged in spaces other than the aurora zone.

[0020] Abnormal surface charging of low-altitude satellites flying in low-altitude orbits 200 km to 1000 km above the Earth is thought to be caused by auroral electrons, and the occurrence of auroras is thought to be caused by the solar wind. A learning device according to the following embodiment creates a learning dataset to be used for machine learning using solar wind data and energy spectrogram data of ions and / or electrons observed by an observation satellite. The learning device then learns using the created learning dataset and constructs a machine learning model. The machine learning model outputs (in other words, predicts) the possibility of charging of an object to be predicted for charging in the auroral zone based on input solar wind data.

[0021] The learning dataset is a dataset of solar wind data and electrostatic charge information generated based on an energy spectrogram of ions and / or electrons. The electrostatic charge information includes information on a surface electrostatic charge label that indicates whether the surface state of the observation satellite is electrostatically charged or not. The inventors discovered that by specifying the conditions used to determine whether an object is electrostatically charged or not, it is possible to construct a machine learning model that can accurately output (predict) the possibility of electrostatic charge on an object to be predicted to be electrostatically charged, without using auroral information, and have proposed the present invention.

[0022] The surface charge labels included in the training dataset are assigned based on the ion and / or electron energy spectrogram data related to the ion flux and / or electron flux observed by the observation satellite. If the surface state of the observation satellite is charged, it is assigned "charged," and if not, it is assigned "uncharged." "Energy spectrogram data relating to ion flux" includes energy-time spectrograms of ion differential energy flux and energy-time spectrograms of ion differential number flux. "Electron flux energy spectrogram data" includes energy-time spectrograms of electron differential energy flux and energy-time spectrograms of electron differential number flux.

[0023] In the first and second embodiments described below, a surface charge label is assigned based on the presence or absence of abnormal values ​​in the electron differential energy flux and the ion differential energy flux. More specifically, in the first embodiment, a surface charge label of “charged” is assigned to energy flux spectrogram data that satisfies all of the conditions (A) to (C) described below, and a surface charge label of “uncharged” is assigned to the rest. In the second embodiment, a surface charge label of “charged” is assigned to energy spectrogram data that satisfies all of the conditions (A) and (D) to (H) described below, and a surface charge label of “uncharged” is assigned to the rest. Condition (A) is a condition related to the electron differential energy flux, and conditions (B) to (H) are conditions related to the presence or absence of abnormal values ​​in the ion differential energy flux.

[0024] In a third embodiment described below, a surface charge label is assigned using an image classification model that classifies an energy-time spectrogram of an ion differential number flux into "charged" or "uncharged." An ion differential particle flux energy-time spectrogram is a visualized diagram showing the ion differential particle flux in two dimensions: energy and time. In the energy-time spectrogram, the temporal change in the energy spectrum is visually shown by a color scale display according to the energy flux value. In the third embodiment, a surface charge label of "charged" is assigned to energy-time spectrogram data that satisfies condition (J) described below, and a surface charge label of "uncharged" is assigned to other data.

[0025] The "object to be predicted to be charged" includes a functioning satellite, a satellite that has exceeded its useful life and has ceased to function, a satellite that has become uncontrollable due to an accident or malfunction, the rocket body or some of its parts used to launch a satellite, debris generated by the separation of a multi-stage rocket, etc. Below, we will take the case where the object to be predicted to be charged is a satellite as an example and explain it as a "prediction target satellite."

[0026] The "aurora zone" is the region where aurora electrons (high-energy electrons) reach low Earth orbit due to magnetospheric activity, and the approximate region is the area between 60 and 80 degrees and between -85 and -55 degrees in the geomagnetic coordinate system (mag_lat (MLAT): geomagnetic latitude). The MLAT area between 60 and 80 degrees is the Arctic aurora zone, and the area between -85 and -55 degrees is the Antarctic aurora zone.

[0027] The above ion and electron energy spectrogram data is acquired by the SSJ (Special Sensor J), a plasma particle measuring instrument mounted on the DMSP (Defense Meteorological Satellite Program) satellite operated by the National Oceanic and Atmospheric Administration (NOAA). The SSJ data acquired by the SSJ includes ion and electron energy spectrogram information related to ion flux and electron flux. On the DMSP satellite, the SSJ opening is grounded on the surface of the satellite. Therefore, when the DMSP satellite's surface is charged to -ΦeV, surrounding ions (below 1 eV) are accelerated by the potential of the DMSP satellite and become ΦeV ions, which pass through the sensor opening. As a result, an extremely high ΦeV ion flux is observed.

[0028] FIG. 3 shows data related to an energy spectrogram of ions and electrons at 18:45:20 (geomagnetic local time) on May 2, 2011 of DMSP-F16 (hereinafter sometimes simply referred to as "energy spectrogram"). The data is information on the fluxes of ions and electrons in an energy range of 30 eV to 30 keV separated into 19 energy channels (Ch1 to Ch19). The 19 energy channels (Ch1 to CH19) have different energy ranges. In FIG. 3, the electron flux 101 in each channel Ch1 to Ch19 is plotted as white circles, and the ion flux 102 is plotted as black circles. In FIG. 3, there are channels (Ch1, Ch4, Ch13, Ch14, Ch16, Ch18) where the ion flux 102 is not plotted, which indicates that ions could not be observed or that they are in trace amounts. In FIG. 3, the vertical axis indicates the energy flux of ions or electrons, and the horizontal axis indicates the energy range assigned to each detection channel. In the example shown in FIG. 3, it can be seen that the ion flux 102 is large and prominent in channel Ch7 with an energy of 300 eV. This means that the DMSP-F16 satellite is surface-charged at -300 eV. The data 9B in FIG. 9 described later is an energy-time spectrogram and corresponds to arranging the data in FIG. 3 in time series.

[0029] In the first embodiment described later, an example of creating a learning dataset using data of an energy spectrogram when the DMSP satellite is present in the auroral zone (60 degrees < MLAT < 80 degrees) in the Arctic region is given. The machine learning model according to the first embodiment outputs, as a likelihood value of 0 to 1, the possibility (or rather, "probability") of charging of the target artificial satellite in the auroral zone in the Arctic region when inputting solar wind parameters (solar wind data) as feature quantities.

[0030] In the second embodiment described below, an example of creating a learning dataset using data of an energy spectrogram when a DMSP satellite is present in the auroral zone on the Antarctic side (-85° < MLAT < -55°) will be given. When the machine learning model according to the second embodiment inputs solar wind parameters as features, it outputs the likelihood of surface charging of the artificial satellite to be predicted in the auroral zone of the Antarctic region with a value from 0 to 1.

[0031] In the third embodiment described below, a learning dataset was created using data of an energy spectrogram when a DMSP satellite was present in the auroral zone of the Arctic region (60° < MLAT < 80°). However, similar to the second embodiment, data of an energy spectrogram when a DMSP satellite is present in the auroral zone on the Antarctic side (-85° < MLAT < -55°) may also be used. When the machine learning model according to the third embodiment inputs solar wind parameters as features, it outputs the likelihood of surface charging of the artificial satellite to be predicted in the auroral zone of the Arctic region with a value from 0 to 1.

[0032] The closer the likelihood is to 1, the higher the possibility of charging, and the closer it is to 0, the lower the possibility of charging. That is, the possibility of surface charging of the artificial satellite to be predicted can be predicted from the value of the likelihood output from the machine learning model.

[0033] Still, a learning dataset may be created using data from the energy spectrogram when the DMSP satellite was present in the auroral zones of the Arctic region (60° < MLAT < 80°) and the Antarctic region (-85° < MLAT < -55°) respectively. And a machine learning model may be constructed using such a learning dataset. The machine learning model thus constructed inputs solar wind parameters as features and outputs the likelihood of surface charging of the target artificial satellite in the auroral zones of the Arctic region and the Antarctic region respectively as values between 0 and 1. Also, when the above machine-leaning model inputs solar wind parameters and the position information (e.g., position information on whether it is located in the auroral zone of the Arctic region or the auroral zone of the Antarctic region) in the geomagnetic coordinate system of the target artificial satellite as features, it outputs the likelihood of surface charging of the target artificial satellite in the auroral zone corresponding to the position information as a value between 0 and 1.

[0034] The above solar wind data is data observed by a solar wind observation satellite. The solar wind data can be obtained, for example, using the OMNIWeb service provided by the Goddard Space Flight Center. The "OMNI data" obtained by the OMNIWeb service is solar wind data. Referring to FIG. 4, the solar wind observation satellite 71 observes the solar wind near the Lagrange point (L1) where the gravity of the sun S and the earth E is balanced. The solar wind reaches the earth E about one hour after leaving the Lagrange point L1. In contrast, the OMNI data observed by the solar wind observation satellite 71 can be obtained almost in real time.

[0035] Hereinafter, each embodiment will be described in detail. In the following description, components similar to those already described may be denoted by the same reference numerals, and the description may be omitted.

[0036] <First Embodiment>

[0037] [Functional Block Configuration of Learning Device]

[0038] 1 shows the configuration of a learning system 10. The learning system 10 includes a learning device 1, which is an information processing device; the Internet N, which is an information communication network; a server 6 that stores and provides data on ion and electron energy spectrograms (SSJ data) 60 observed by the DMSP satellite; and a server 7 that stores and provides solar wind data (OMNI data) 70. The learning device 1 includes a communication unit 2, a control unit 3, a memory unit 4, and a data input unit 8.

[0039] The communication unit 2 is configured to be able to communicate with various servers 6 and 7 via the Internet N. The data input unit 8 acquires energy spectrogram data 60 and solar wind data 70 from the servers 6 and 7 via the communication unit 2.

[0040] The storage unit 4 includes a machine learning model 40 and a training dataset storage unit 41 used when constructing the machine learning model 40. Details of the training dataset will be described later.

[0041] The control unit 3 creates a learning dataset by associating data 9B of the energy-time spectrogram of the differential energy flux of ions and electrons, as shown in FIG. 9 , acquired by the data input unit 8 with solar wind data, and stores the created learning dataset in the learning dataset memory unit 41.

[0042] In addition, the control unit 3 uses the learning dataset stored in the learning dataset memory unit 41 to train the machine learning model 40, and when solar wind parameters (solar wind data) are input as features, constructs the machine learning model 40 which outputs the possibility (probability) of charging of the target artificial satellite in the auroral zone of the Arctic region as a likelihood value between 0 and 1.

[0043] In this embodiment, an example has been given in which one information processing device (learning device) creates a learning dataset and builds a machine learning model, but the information processing device that creates the learning dataset and the information processing device (learning device) that builds the machine learning model may be separate devices.

[0044] The control unit 3 can be realized by, for example, a processor such as a CPU (Central Processing Unit), a working memory, and a non-volatile storage device. A control program to be executed by the processor is stored in the storage device, and the processor loads the program into the working memory and executes it, thereby fulfilling the functions of the control unit 3. The storage device may be part of the memory unit 4. In the learning system 10, the control program causes the learning device 1 to execute processes including the steps shown in Figures 7 and 8.

[0045] [Hardware configuration of the learning device]

[0046] FIG. 2 is a diagram showing the hardware configuration of the learning device 1.

[0047] 2, the learning device 1 includes hardware necessary for configuring a computer, such as a CPU 51, a ROM (Read Only Memory) 52, a RAM (Random Access Memory) 53, an input / output interface 55, a bus 54 connecting these components together, and a storage device such as a HDD. For example, the learning device 1 can be realized by any computer, such as a PC (Personal Computer).

[0048] The CPU 51 accesses the RAM 53 and other memory as needed, performs various arithmetic processing, and overall controls each block of the learning device 1. The ROM 52 is a non-volatile memory that permanently stores firmware such as the OS program and various parameters to be executed by the CPU 51. The RAM 53 is used as a working area for the CPU 51, and temporarily stores the OS, various applications currently being executed, various data currently being processed, and the like.

[0049] For example, the learning method according to this embodiment is executed by the CPU 51 loading a learning program according to this embodiment, which is stored in advance in the ROM 52 or the like, into the RAM 53 and executing it. The CPU 51 or the like executes a predetermined program, thereby configuring the control unit 3 as a functional block. The program is installed in the learning device 1, for example, via various recording media. Alternatively, the program may be installed via the Internet or the like. There are no limitations on the type of recording medium on which the program is stored, and any computer-readable recording medium may be used. For example, any computer-readable non-transitory recording medium may be used.

[0050] The input / output interface 55 is connected to a display unit 56, an operation reception unit 57, a storage unit 58, a communication unit 59, and the like.

[0051] The display unit 56 is a display device that uses, for example, an LCD (Liquid Crystal Display), an OLED (Organic Electro Luminescence Display), or the like.

[0052] The operation reception unit 57 is, for example, a pointing device such as a mouse, a keyboard, a touch panel, or other input device. When the operation reception unit 57 is a touch panel, the touch panel can be integrated with the display unit 56.

[0053] The storage unit 58 is, for example, a non-volatile memory such as a hard disk drive (HDD), a flash memory (solid state drive (SSD)), or other solid-state memory.

[0054] The communication unit 59 is, for example, a NIC (Network Interface Card) for Ethernet or various modules for wireless communication such as wireless LAN, and is responsible for communication processing between the server where SSJ data is stored and the server where OMNI data is stored.

[0055] Although not shown, the basic hardware configuration of the information processing apparatus 110 to be described later is substantially the same as the hardware configuration of the learning apparatus 1.

[0056] [Learning dataset]

[0057] The learning dataset used for machine learning is a dataset in which solar wind data (OMNI data) and charging information are associated with each other. The charging information is created based on data (SSJ data) of the energy spectrogram of the DMSP satellite, which is an observation artificial satellite.

[0058] In the present embodiment, using the SSJ data of DMSP-F13 to DMSP-F18, the data from 1996 to December 31, 2021 is used as learning / verification data, and the data from January 1, 2022 to December 31, 2022 is used as inference data.

[0059] More specifically, a learning dataset is created by associating the charging information based on the SSJ data with MLAT (Magnetic Latitude) in the geomagnetic coordinate system of 60 degrees < MLAT < 80 degrees, and MLT (Magnetic Local Time) in the geomagnetic coordinate system of 0 to 6 hours and 20 to 24 hours from 1996 to December 31, 2021, and the OMNI data, and machine learning of the machine learning model 40 is performed using the learning dataset.

[0060] Also, in the verification of the learned machine learning model to be described later, the SSJ data from January 1, 2022 to December 31, 2022 is used as inference data, and the performance of the classifier is confirmed by AUC (Area Under the ROC Curve).

[0061] Table 1 shows the charging information created based on the data of the energy spectrogram observed by the DMSP satellite (specifically, the energy-time spectrogram of electron differential energy flux and ion differential energy flux). The charging information is used when creating the learning dataset.

[0062] [Table 1]

[0063] Table 2 shows the solar wind data used to create the learning dataset.

[0064] [Table 2]

[0065] Table 3 shows the contents of the learning data set created by the learning device 1. Table 3 also serves as an explanation of the terms used in Tables 1 and 2.

[0066] [Table 3]

[0067] In creating the training dataset, the solar wind data and the electrification information were associated based on the time (date) that the solar wind reached the Earth. In other words, the training dataset was created by associating solar data and electrification information on the same date.

[0068] As shown in Table 1, the charge information includes information related to the latitude in the geomagnetic coordinate system (MLAT), the local time in the geomagnetic coordinate system (MLT), the surface charge label (Charge label), the surface charge potential (Charge Value), and the time (date) when the solar wind reached the Earth.

[0069] The surface charge label information is information on whether the surface state of the DMSP satellite is "charged" or "uncharged." When the energy spectrogram data satisfies all of the following conditions (A) to (C), the surface charge label "charged" is assigned to the energy spectrogram data, and otherwise the surface charge label "uncharged" is assigned. In this way, the surface charge label is expressed as a binary value. For example, "charged" may be expressed as "1" and "uncharged" as "0."

[0070] Conditions (A) to (C) used to determine whether to assign a surface charge label of "charged" or "uncharged" to energy spectrogram data will be described.

[0071] (A) In the energy spectrogram data, the differential energy flux of electrons above 14 keV is 10 8 eV / (cm 2 ·ΔeV·ster·s) or more.

[0072] (B) In the energy spectrogram data, there is an abnormal value of the ion differential energy flux in the energy range from 95 eV to 2040 eV.

[0073] (C) The anomaly of the differential energy flux of ions observed by the DMSP satellite in the energy spectrogram data is 5 × 10 6 eV / (cm 2 ·ΔeV·ster·s) or more.

[0074] The above-mentioned "outliers" are detected using the Smirnoff-Grubbs test, an outlier test.

[0075] Condition (A) above means that when the DMSP satellite passes through the auroral zone, it observes the precipitating electrons that cause aurorae.

[0076] The above condition (B) means that the DMSP satellite has a surface charge. For example, in a satellite mission involving docking operations, the surface charge value at which accidents such as damage to systems and equipment are likely to occur is -100V or less, and in the worst case scenario, -1400V. Therefore, the range is limited to 95eV or more and 2040eV or less.

[0077] The condition (C) above was set to remove abnormal values ​​that are too small. The condition (C) above specifies a specific abnormal value, and further limits the range that satisfies the condition (B) above. The state in which there is one point where the value of the ion differential energy flux is prominent as shown in FIG. 3 indicates that there is an abnormal value regarding the ion differential energy flux. In this embodiment, the abnormal value is set to 5×10 6 eV / (cm 2 The condition (C) is set as follows:

[0078] Note that the SSJ data from the DMSP satellite is data obtained at one-second intervals, whereas the OMNI data is data obtained at one-minute intervals. In this embodiment, for example, if an energy spectrogram containing 60 pieces of data (one minute's worth) acquired every second between 10:00:00 and 10:00:59 on M / D / Y / Y has at least one location that satisfies all of the above conditions (A) to (C), a surface charge label of "charged" is assigned to the data at 10:00:00 on M / D / Y / Y, and charge information (see Table 1) is generated. On the other hand, if no location satisfies all of the above conditions (A) to (C) among the one-minute worth of energy spectrogram data, a surface charge label of "uncharged" is assigned to the data at 10:00:00 on M / D / Y / Y, and charge information is generated. Then, the electrostatic charge information for the date Y / M / D / 10:00:00 is associated with the OMNI data for the date Y / M / D / 10:00 (see Table 2), and a training dataset is created. In other words, if an energy spectrogram is observed in which a location that satisfies all of the above conditions (A) to (C) appears at least once within one minute, electrostatic charge information including the surface electrostatic charge label "charged" is associated with the corresponding OMNI data. In this way, the training dataset is a plurality of datasets in which solar wind data and electrostatic charge information are associated with each other for each minute.

[0079] Figure 5 compares the distribution of the AE INDEX (auroral index) when a target is charged and when it is not charged, with data that shows a charge at least once within one minute being considered charged (charge) and data that shows no charge being considered no charge (no charge). In Figure 5, the solid line represents uncharged data 103, and the dotted line represents charged data 104. Note that the uncharged data 103 is limited to magnetic latitudes in the auroral belt only. As shown in Figure 5, the distribution of the auroral index differs when the target is charged and when it is not charged, and it can be said that the auroral index can be used to detect whether or not a target is charged.

[0080] As shown in Table 2, the solar wind data (OMNI data) includes information on the x-component (Vx) of the solar wind velocity, the y-component (Vy) of the solar wind velocity, the z-component (Vz) of the solar wind velocity, the x-component (Bx) of the solar wind magnetic field, the y-component (By) of the solar wind magnetic field, the z-component (Bz) of the solar wind magnetic field, the solar wind pressure (Pressure), the solar wind temperature (Temperature), the solar wind proton density (proton density), the ratio (beta) of the plasma pressure to the magnetic pressure, and the time (date) when the solar wind reaches the Earth.

[0081] In this embodiment, when any solar wind reaches the Earth, a machine learning model for binary classification of whether a target artificial satellite existing at any location within the auroral zone "surface-charges" or "does not surface-charge" is constructed using the above learning dataset. When the constructed machine learning model (trained machine learning model) inputs solar wind parameters (solar wind data) as features, it outputs the likelihood of the target artificial satellite charging within the auroral zone (60° < MLAT < 80°, MLT is 0 to 6 hours, 20 to 24 hours) as a value between 0 and 1.

[0082] (Preprocessing of the learning dataset)

[0083] It is preferable to perform preprocessing of the learning dataset for the purpose of appropriately organizing the learning dataset and constructing a high-quality learning model.

[0084] Specifically, in the creation of the learning dataset, preprocessing of the learning dataset was performed so that only data when the DMSP satellite was within the auroral zone was adopted. In this embodiment, as described above, the data used was limited to 60° < MLAT < 80°, MLT is 0 to 6 hours, 20 to 24 hours.

[0085] This embodiment is based on the assumption that surface charging of low-earth-orbit satellites (e.g., DMSP satellites) is caused by auroral electrons, and that auroras occur in response to changes in the solar wind. Therefore, when the DMSP satellite is not in the auroral zone, it does not depend on solar wind parameters, and surface charging of the DMSP satellite does not occur. For this reason, only data from when the DMSP satellite is in the auroral zone was used, as described above.

[0086] Furthermore, the number of data (samples) assigned the surface charge label "uncharged" is clearly greater than the number of data (samples) assigned the surface charge label "charged," resulting in an unbalanced training data set. In contrast, in this embodiment, preprocessing was performed to undersample the data assigned "uncharged" so that the ratio between the number of data assigned "charged" and the number of data assigned "uncharged" was approximately 1:1.

[0087] Furthermore, if undersampling is performed uniformly with respect to the satellite's position in the geomagnetic coordinate system, the distribution of positions will differ between data assigned with "charged" and data assigned with "uncharged." If the distribution of positions differs, this could result in a simplistic model in which good accuracy can be achieved by simply relying on the ratio of data assigned with "charged" to data assigned with "uncharged" for each position in the training dataset. Therefore, in this embodiment, the auroral zone is divided into 1-degree increments in the latitude direction and 1-hour increments in the local time direction, and undersampling is performed on the data assigned with "uncharged" so that the ratio of the number of data assigned with "charged" to the number of data assigned with "uncharged" within each divided region is approximately 1:1.

[0088] Figure 6(A) shows the distribution of the number of charged or uncharged counts at geomagnetic latitudes in an undersampled training data set. Figure 6(B) shows the distribution of the number of charged or uncharged counts at local time in an undersampled training data set. In Figures 6(A) and 6(B), three types of dots with different densities are used, with the distribution 105 of the number of counts as "no charge" shown with coarse dots, the distribution 106 of the number of counts as "charge" shown with the next coarser dots, and the overlapping portion 107 of the distribution of the number of counts as "charged" and the distribution of the number of counts as "uncharged" shown with the densest dots.

[0089] As shown in Figures 6(A) and (B), in the undersampled training dataset, the ratio of the number of data assigned "charged" to the number of data assigned "uncharged" is approximately 1:1 within regions divided into 1-degree increments in the latitude direction and 1-hour increments in the local time direction. In this way, by creating a training dataset that matches the distribution of DMSP satellite position information when "charged" is assigned and when "uncharged" is assigned, it is possible to avoid incorrect classification due to bias in position information.

[0090] In this way, by preprocessing the training dataset, the training dataset can be made appropriate and a high-quality classification model can be constructed.

[0091] [How to create a training dataset]

[0092] A method for creating a learning dataset by the control unit 3 of the learning device 1 will be described according to the flow in Figure 7. In the specification, "ST" means "step." The same applies to Figures 7, 8, 12, and 19.

[0093] The control unit 3 of the learning device 1 acquires energy spectrogram data 60 observed by the DMSP satellite (ST1).

[0094] Next, based on the acquired energy spectrogram, if there is at least one piece of data among the energy spectrograms containing data for one minute that satisfies all of the above conditions (A) to (C), the control unit 3 assigns a surface charge label of “charged,” and assigns a surface charge label of “uncharged” to the rest, and generates charge information shown in Table 1 that includes information related to the surface charge label (ST2).

[0095] Next, the control unit 3 acquires the solar wind data 70 shown in Table 2 (ST3).

[0096] Next, the control unit 3 uses the generated electrification information and the acquired solar wind data to create a learning data set so that electrification information and solar wind data with the same date are paired (ST4). The control unit 3 stores the created learning data in the learning data set storage unit 41.

[0097] [Learning Method]

[0098] A method for learning a machine learning model in the learning device 1 will now be described.

[0099] Because the training dataset is structured data in a tabular format and model interpretability is required, XGBoost (Extreme Gradient Boosting) was adopted as the machine learning algorithm used to train the machine learning model 40. The interpretability of the above model refers to understanding the reasons and causes behind the predictions made by the machine learning model. A model with high interpretability can clearly show how the prediction result was obtained and which features significantly contributed to the prediction. XGBoost is an algorithm that combines multiple weak learners (usually decision trees) to build a single powerful model. Each learner is trained to correct errors made by the previous learner.

[0100] The characteristics of XGBoost are as follows:

[0101] Gradient Boosting: XGBoost is an algorithm based on the gradient boosting framework, in which weak learners (often decision trees) are trained sequentially and their predictions are combined to produce a powerful model.

[0102] High interpretability: XGBoost is said to have high model interpretability because it makes it easy to understand feature importance, which is especially useful for tabular data.

[0103] Regularization and overfitting prevention: XGBoost includes a regularization term to reduce the risk of overfitting.

[0104] Efficiency and scalability: XGBoost is computationally efficient and can be applied to large datasets.

[0105] Table 4 explains the hyperparameters of XGBoost. XGBoost has hyperparameters, which were tuned using Bayesian optimization.

[0106] [Table 4]

[0107] The learning method will be explained using FIG. 8. FIG. 8 shows the flow of the learning method when constructing a machine learning model. As shown in FIG. 8, the control unit 3 acquires a learning dataset stored in the learning dataset storage unit 41 (ST11). Next, the control unit 3 uses the learning dataset to train the machine learning model 40 (ST12). By repeating this process, the machine learning model 40 is constructed.

[0108] [Verification of trained model]

[0109] We verified a trained machine learning model using a training dataset that correlated electrification information based on SSJ data from DMSP-F13 to DMSP-F18 with solar wind data (OMNI data) from 1996 to 2021. In the verification, the DMSP-F18 data from January 3, 2022 was used as inference data.

[0110] Result 9A in the upper part of Figure 9 shows the training dataset (true label) and the machine learning model output (logit) for the ion and electron energy spectrograms acquired by the DMSP-F18 satellite on January 3, 2022. Data 9B in the lower part of Figure 9, with energy [eV] on the vertical axis and date on the horizontal axis, shows the energy-time spectrogram of the ion differential energy flux acquired by the DMSP-F18 satellite and the energy-time spectrogram of the electron differential energy flux acquired by the DMSP-F18 satellite. In Data 9B (energy-time spectrogram), the vertical axis indicates the energy range assigned to the detection channel, the horizontal axis indicates time, and the energy flux values ​​are represented by color. In Figure 9, Result 9A is displayed on the same time axis as Data 9B.

[0111] In Result 9A of Figure 9, the logit at the top shows the output of a trained machine learning model that was input with solar wind data from January 3, 2022. The trained machine learning model outputs the possibility of charging as a likelihood between 0 and 1. In the figure, the higher the likelihood value, the higher the dot density. In Result 9A, the true label at the bottom shows the result of assigning "charged" to data that meets all of the above conditions (A) to (C) based on the energy spectrogram data acquired by the DMSP-F18 satellite on January 3, 2022, and assigning "uncharged" to all other data. The true label is expressed as two values: "charged" and "uncharged." In the figure, "charged" is shown with a higher dot density than "uncharged." Because the solar wind data used was data at one-minute intervals, the surface charge labels were created at one-minute intervals.

[0112] As shown in Figure 9A, the true label assigned "uncharged" to the data at 10:06 and "charged" to the data at 10:07. In contrast, the trained model output a likelihood of approximately 0.5 for the 10:06 data and approximately 0.8 for the 10:07 and 10:08 data in logit. When the likelihood threshold was set to 0.5, all of these events were considered surface-charged events. Figure 10 shows the confusion matrix used to verify the inference accuracy of the trained model using validation data with a threshold of 0.5. The accuracy, recall, precision, F1 score, and AUC are shown in Table 5. The accuracy, recall, precision, F1 score, and AUC are evaluation metrics used to evaluate the performance of a model (classifier) ​​after it has been trained using machine learning for classification problems. In addition, since the number of positive and negative examples in the inference data is not the same, the accuracy rate, recall rate, precision rate, and F1 score are left blank in Table 5.

[0113] Table 5 shows the results for the validation data and inference data when the threshold is set to 0.5.

[0114]

Table 5

[0115] As shown in Table 5, the AUC score was 0.833 for the validation data and 0.869 for the inference data, confirming that the created machine learning model was very good. Also, as shown in Table 5, the results of the confusion matrix had a good balance with similar accuracy, recall, and f1-score values.

[0116] [Information Processing System Using a Trained Learning Model]

[0117] An example of an information processing system using the above-mentioned trained machine learning model (hereinafter sometimes referred to as "trained model") will be given. When OMNI data of a certain date and time is input, the above-mentioned trained model outputs the possibility of charging of a target artificial satellite in the auroral zone of the Arctic region (more specifically, 60 degrees < MLAT < 80 degrees, MLT is 0 to 6 hours, 20 to 24 hours) caused by the solar wind reaching the Earth about 1 hour after that date and time. Specifically, the trained model outputs the possibility of charging of a target artificial satellite in a low-altitude orbit about 830 km to 850 km from the ground, which is the flight altitude of the DMSP satellite. The trained model of the present embodiment outputs the possibility of charging from 830 km to 850 km from the ground, which is the flight altitude of the DMSP satellite.

[0118] FIG. 11 is an example of an information processing system 100. As shown in FIG. 11, the information processing system 100 includes an information processing device 110, the Internet N, a server 7 that stores and provides solar wind data 70, and one or more user terminals 20.

[0119] (Information Processing Device)

[0120] The information processing device 110 is, for example, a web server operated by the operator of a predicted electrification information providing website. The information processing device 110 is connected via the Internet N to one or more user terminals 20 and a server 7 that provides solar wind data (OMNI data) 70. The information processing device 110 provides a service of providing predicted electrification information to users of the user terminals 20.

[0121] In this embodiment, the information processing device 110 uses the possibility of charging (probability of charging, expressed as a value between 0 and 1) output by the trained model 150 to present to the user, as predicted charging information, a display image 123 (see FIG. 13 ) that visualizes, as a distribution map, an area in the auroral zone of the Arctic region where a prediction target satellite flying in a low-altitude orbit is likely to be charged at an arbitrary date and time (for example, specifically, a date and time specified by a user's input operation). In FIG. 13 , the higher the probability of charging, the denser the dots. The above-mentioned "low-altitude orbit" specifically refers to the altitude range of 830 km to 850 km of the DMSP satellite that observed the ion and electron energy spectrogram data used in the training dataset. Note that in this embodiment, as shown in FIG. 13 , an example has been given in which the display format of the predicted charging information presented to the user is a distribution map, but the display format is not limited to this.

[0122] As shown in FIG. 11, the information processing device 110 includes a communication unit 11, a control unit 12, and a storage unit 15.

[0123] The communication unit 11 is configured to be able to communicate with the server 7 and one or more user terminals 20 via the Internet N. The communication unit 11 acquires solar wind data (OMNI data) 70 from the server 7. The communication unit 11 acquires input operation information by a user on the user terminal 20. Specifically, the communication unit 11 acquires date and time information as input operation information from the user terminal 20. The date and time information is, for example, an arbitrary date and time set by the user, and is input in minutes. In addition, the communication unit 11 provides predicted electrification information to the user terminal 20 via the Internet N.

[0124] The storage unit 15 stores a trained model (trained machine learning model) 150.

[0125] The control unit 12 includes a calculation unit 121 and a display image generation unit 122 .

[0126] The calculation unit 121 acquires from the server 7 solar wind data 70 for an arbitrary date and time (date and time information) set by the user obtained from the user terminal 20. The solar wind data 70 acquired here is data observed at the arbitrary date and time set by the user. Specifically, the solar wind data 70 includes information related to the x-component (Vx) of the solar wind speed, the y-component (Vy) of the solar wind speed, the z-component (Vz) of the solar wind speed, the x-component (Bx) of the solar wind magnetic field, the y-component (By) of the solar wind magnetic field, the z-component (Bz) of the solar wind magnetic field, the solar wind pressure, the solar wind temperature, the solar wind proton density, and the ratio of plasma pressure to magnetic pressure (beta).

[0127] Furthermore, the calculation unit 121 calculates the possibility of charging by inputting solar wind data using the trained model 150. The trained model 150 outputs the possibility of charging of the target satellite flying in a low-altitude orbit in the auroral zone of the Arctic region, caused by the solar wind that will reach the Earth approximately one hour after the above-mentioned arbitrary time, as a numerical value between 0 and 1.

[0128] The display image generation unit 122 generates a display image (predicted charging information) 123 to be transmitted to the user terminal 20 by using the information on the possibility of charging calculated by the calculation unit 121. An example of the display image 123 is shown in FIG. 13. The display image 123 is a distribution map that visualizes the possibility of charging of the predicted target satellite in the region where 60 degrees < MLAT < 80 degrees, MLT is from 0 to 6 hours and from 20 to 24 hours, when there is a predicted target satellite in the region. The display image 123 can be an image whose color changes according to the likelihood value (probability of charging). Further, for example, a display image can be generated that visualizes so that the user can recognize that there is a possibility of "charging" at a position where the likelihood is 0.5 or more. The example shown in FIG. 13 maps and visualizes the possibility of charging in the auroral zone when looking at the Earth, and represents the difference in the possibility of charging by varying the coarseness density of dots. In the example shown in FIG. 13, the closer the dot density is to being dense, the closer the output value (likelihood) is to 1, indicating a high possibility of charging, and the coarser the dot density is, the closer the output value (likelihood) is to 0, indicating a low possibility of charging. By looking at the display image shown in FIG. 13, the user can intuitively recognize the regions with high and low possibilities of charging.

[0129] As described above, the information on the possibility of charging output from the learned model 150 indicates the possibility of charging of a predicted target satellite flying in a low altitude orbit (altitude 830 km to 850 km in this embodiment) in the auroral zone of the Arctic region, caused by the solar wind that reaches the Earth about one hour after any of the above times. Therefore, by grasping the predicted charging information in the future auroral zone, the user can select a position and timing (time) with a low possibility of charging and execute, for example, docking operation or the like. Thereby, the risk of damage to the system and equipment of the satellite due to discharge caused by the potential difference between the satellites can be suppressed.

[0130] The control unit 12 can be realized by, for example, a processor such as a CPU, a working memory, and a non-volatile storage device. A control program executed by the processor is stored in the storage device, and the processor reads the program into the working memory and executes it, thereby fulfilling the functions of the control unit 12. The storage device may be part of the storage unit 15. In the information processing system 100, the control program causes the information processing device 110 to execute a series of processes for generating predicted electrification information (display image 123 in this embodiment) to be presented to the user, including a solar wind data acquisition step, an electrification possibility calculation step, an electrification prediction step, and a display image generation step.

[0131] (user terminal)

[0132] The user terminal 20 is a terminal used by a user, and is, for example, a smartphone, a mobile phone, a tablet PC (Personal Computer), a notebook PC, a desktop PC, etc. The user terminal 20 includes a communication unit 21, an input unit 22, and a display unit 23.

[0133] The communication unit 21 is configured to be able to communicate with the information processing device 110 via the Internet N. The communication unit 21 transmits information such as input operation information by a user performed on the input unit 22 of the user terminal 20 to the information processing device 110. The communication unit 21 receives predicted electrification information (display image, etc.) generated by the information processing device 110 from the information processing device 110.

[0134] The display unit 23 displays the received display image, etc. The display unit 23 is a display device using, for example, a liquid crystal display, an organic electroluminescence (EL) display, etc. The input unit 22 is an operation device such as a keyboard or a mouse for inputting user operations. The input unit 22 may be a touch panel, etc., which is configured integrally with the display unit 23.

[0135] (others)

[0136] Figure 14 shows a histogram of the model output for each surface charge label ("charged" and "uncharged") in the validation data. Figure 14 uses three types of dots with different densities, with the distribution of "no charge" numbers 108 shown with sparse dots, the distribution of "charge" numbers 109 shown with the next sparser dots, and the overlapping area 111 between the distribution of "charged" and the distribution of "uncharged" with the densest dots. As shown in Figure 14, when the output of the trained model is between 0.4 and 0.6, the number of "uncharged" and "charged" samples is almost the same. However, when the output of the trained model is below 0.4 or above 0.6, there is a gap between the number of samples. In particular, it can be seen that almost no misclassification occurs when the output of the trained model is below 0.2 or above 0.8. Figure 15 shows the confusion matrix used when verifying the inference accuracy of a trained model using validation data, with the output of the trained model 150 limited to samples below 0.2 or above 0.8, and with a threshold of 0.5. Comparing Figure 10 with Figure 15, it can be seen that the number of misclassified samples is significantly smaller when the output of the trained model 150 is below 0.2 or above 0.8.

[0137] In the above, an example was given in which the control unit 12 of the information processing device 110 generates a distribution map of the possibility of electrification throughout the auroral zone using likelihood information of 0 to 1 output from the trained model 150, but this is not limited to this, and predicted electrification information to be presented to the user may be generated depending on the purpose.

[0138] For example, when using predicted electrification information in a docking operation between a new satellite and an already-in-orbit satellite for the purposes of space debris removal, on-orbit servicing, etc., it is more important to detect when surface charging will not occur than when surface charging will occur in order to avoid discharge accidents caused by surface charging. Thus, when a user uses predicted electrification information generated by the information processing device 110 for the purpose of avoiding discharge accidents, the accuracy of the prediction of "no charge" is high when the output is 0.2 or less, as described above. Therefore, the predicted electrification information presented to the user may suggest a docking operation in the aurora zone only when the output of the trained model 150 is 0.2 or less.

[0139] As another example, if a user requires predicted electrostatic charge information at a specific location in the auroral zone at a specific date and time, the control unit 12 may acquire, from the user terminal 20, in addition to date and time information, location information of the location for which the user requires predicted electrostatic charge information (location information of the specific location). The control unit 12 acquires solar wind data based on the date and time information acquired from the user terminal 20, and calculates information on the possibility of electrostatic charge using the trained model 150. The control unit 12 may then use the information on the possibility of electrostatic charge and the location information acquired from the user terminal 20 to generate predicted electrostatic charge information indicating the possibility of electrostatic charge at the specific location, and present this to the user.

[0140] As yet another example, if a user requires predicted electrostatic charge information for a specific artificial satellite at a specific date and time, the control unit 12 may acquire, from the user terminal 20, information about the specific artificial satellite desired by the user in addition to date and time information. The information about the specific artificial satellite may be, for example, the name of the artificial satellite. The control unit 12 acquires solar wind data based on the date and time information acquired from the user terminal 20 and calculates the possibility of electrostatic charge using the trained model 150. Furthermore, the control unit 12 may calculate position information about the location of the specific artificial satellite at a specific date and time using information from an existing database that stores names of various artificial satellites (including space debris), orbital information, velocity information, position information, and the like associated with the names. The control unit 12 may then generate predicted electrostatic charge information to present to the user using the calculated position information of the specific artificial satellite at a specific date and time and information about the possibility of electrostatic charge for the specific artificial satellite.

[0141] [Information processing method using trained models]

[0142] FIG. 12 is a flow chart of an information processing method using the trained model 150, which is executed by the control unit 12 of the information processing device 110.

[0143] As shown in FIG. 12, the control unit 12 acquires solar wind data 70 for an arbitrary date and time from the server 7 based on date and time information input on the user terminal 20 (an arbitrary date and time set by the user; input operation information) (ST21).

[0144] Next, the control unit 12 uses the trained model 150 to calculate the possibility of electrification from the acquired solar wind data (ST22).

[0145] Next, the control unit 12 generates a display image (predicted charge information) shown in Fig. 13 using the calculated information on the possibility of charge (ST23). The generated display image is transmitted to the user terminal 20 and displayed on the display unit 23 of the user terminal 20.

[0146] In this embodiment, date and time information is input from the user terminal 20, but it is also possible to input the latest solar wind data in ST21 and generate the latest predicted electrification information together with the predicted time in ST23. By doing so, events such as docking and debris retrieval scheduled in an area where electrification is possible at the time generated in ST23 can be canceled, preventing damage to equipment.

[0147] Second Embodiment

[0148] In the learning device according to the second embodiment, the conditions for determining whether a surface is "charged" or "uncharged" are more limited than in the first embodiment when assigning a surface charge label included in the charge information of the learning data used to construct a learning model. The configurations of the learning device and the information processing device using a trained machine learning model are the same as in the first embodiment. Furthermore, the machine learning algorithm used to construct (learn) the machine learning model is XGBoost, the same as in the first embodiment. The following description will focus on the "learning dataset," which is a difference from the first embodiment, and will omit a description of the configuration that is the same as in the first embodiment.

[0149] [Training dataset]

[0150] In this embodiment, when assigning the surface charge label included in the charge information used in the learning dataset, if the above condition (A) and the conditions (D) to (H) described below are all satisfied, the surface charge label “charged” is assigned, and otherwise the surface charge label “uncharged” is assigned.

[0151] (D) In ​​the energy spectrogram data, there are three or fewer outliers for the differential energy flux of ions in all channels.

[0152] (E) The largest ion differential energy flux anomaly in the energy spectrogram data is more than twice the second largest anomaly.

[0153] (F) The largest ion differential energy flux anomaly in the energy spectrogram data is in the energy range from 95 eV to 2040 eV.

[0154] (G) The largest ion differential energy flux anomaly in the energy spectrogram data is 5×10 6 eV / (cm 2 ·ΔeV·ster·s) or more than 10 10 eV / (cm 2 ·ΔeV·ster·s).

[0155] (H) Energy spectrogram data that meets all of the conditions (A), (D) to (G) is observed twice or more consecutively (for 2 seconds).

[0156] The above conditions (D) to (F) are more restrictive than the above condition (B) and are conditions for suppressing noise. The condition (D) means that when plotting the ion flux values ​​of each channel in the energy spectrogram (see Figure 3) to draw a line graph, data (samples) that show two or more peaks in the line graph must be excluded.

[0157] The above condition (G) is a more restrictive version of the above condition (C). The condition (G) is set in order to remove abnormal values ​​that are too small and too large.

[0158] The above condition (H) is set for the purpose of removing noise data, since sudden abnormal values ​​are likely to be noise. Note that condition (H) does not necessarily have to be present, and a surface charge label of "charged" may be assigned to energy spectrogram data that satisfies all of conditions (A) and (D) to (G), and "uncharged" may be assigned to other data. However, from the perspective of enabling more accurate charge prediction, it is preferable to include condition (H), as in this embodiment.

[0159] Here, the higher concept of the above condition (B) is (B)´ The condition that there is an abnormal value for the ion flux. Set The above condition (C) was conceptualized. (C)´ The condition that there is one point where the ion energy flux protrudes as shown in Figure 3. Set.

[0160] It has been found that if the above conditions (B)' and (C)' are met, and condition (B)' is met for 22 seconds or more consecutively, the DMSP satellite is in a surface-charged state. It is important to detect condition (C)' with high accuracy, and in this embodiment, the Smirnoff-Grubbs test is used, and the above conditions (D) to (G) are set as conditions for detecting data that meets condition (C)' with high accuracy. Data that meets conditions (B)' and (C)', and condition (B)' is met for 22 seconds or more consecutively, is nearly the same (but not exactly the same) as data that meets all of the above conditions (D) to (G).

[0161] Data that satisfies each of the conditions (E) to (G) is included in the data that satisfies the condition (D). For this reason, the conditions used when assigning charge labels may not include the conditions (E) to (H), and a surface charge label of "charged" may be assigned to energy spectrogram data that satisfies all of the conditions (A) and (D), and "uncharged" may be assigned to the rest. However, from the perspective of enabling more accurate charge prediction, it is preferable to include the conditions (E) to (H), as in this embodiment.

[0162] Furthermore, data that satisfy both of the above conditions (B)' and (C)' is almost always included in the data that satisfy the above condition (A). Therefore, in the first and second embodiments, the conditions used when attaching the charge labels do not need to include the above condition (A). The conditions used when attaching the charge labels may be (B) and (C) (a modified version of the first embodiment), or (D) to (H) or (D) to (G) (a modified version of the second embodiment). Not including the condition (A) means that only the conditions related to the presence or absence of anomalous values ​​in the ion flux are used. In this case, the surface charge labels are attached based on energy spectrogram data related to the ion flux. However, from the viewpoint of enabling more accurate charge prediction, it is preferable to include the condition (A). Furthermore, the condition used when attaching the charge label may be only (A), but from the viewpoint of enabling more accurate charge prediction, it is preferable to use conditions (A) to (C) (first embodiment) or conditions (A) and (D) to (H) (second embodiment). Note that using only condition (A) means using only conditions related to electron flux. In this case, the surface charge label is attached based on energy spectrogram data related to electron flux.

[0163] As described above, in the first and second embodiments, examples have been given in which a surface charge label is assigned based on the presence or absence of an abnormal value in the electron differential energy flux and the presence or absence of an abnormal value in the ion differential energy flux in the energy spectrogram data relating to the ion differential energy flux and the electron differential energy flux. However, a surface charge label may also be assigned based on the presence or absence of an abnormal value in the ion differential energy flux or the electron differential energy flux.

[0164] As another modification of the first embodiment, the conditions used when applying the charged label may be (A) and (B) (without (C)), (A) and (C) (without (B)), (B) only (without (A) and (C)), or (C) only (without (A) and (B)). However, from the viewpoint of enabling more accurate charge prediction, it is preferable to include conditions (A) to (C).

[0165] As another modification of the second embodiment, the conditions used when applying the charged labels may include (B) and / or (C) in addition to (A), (D) to (H), thereby enabling more accurate charge prediction. As another modification of the second embodiment, the conditions used when applying the charged labels may be (D) only, (A), (B) and (D), (A), (B), (D) to (G), (A) to (G), (A) to (H), (B) to (G), (B) to (H), or (A) to (D).

[0166] In this embodiment, as in the first embodiment, the learning data set used for machine learning of the machine learning model 40 is one in which solar wind data and electrostatic charge information are associated with each other. In the second embodiment, the electrostatic charge information is created based on SSJ data from the DMSP satellite, as in the first embodiment, but the conditions for assigning the surface electrostatic charge label "electrostatic charge" are more limited than in the first embodiment. As a result, as will be described later, it is possible to obtain a trained model that can detect surface electrostatic charge that is more dangerous to satellites with better classification performance. Therefore, it is possible to perform electrostatic charge predictions using this trained model with higher accuracy.

[0167] Furthermore, in the first embodiment, the charged labels were assigned with a resolution of one minute, whereas in the second embodiment, the resolution was increased to one second to create a learning data set.

[0168] The SSJ data of the DMSP satellite is data at 1-second intervals, while the OMNI data is data at 1-minute intervals. In the present embodiment, charge information is generated for each SSJ data acquired every second. Then, for each of the data of the energy spectrogram including 60 pieces (for 1 minute) of data acquired every second between 10:00:00 and 10:00:59 on Y year M month D day, the OMNI data with date being 10:00 on Y year M month D day (see Table 2.) is associated with the generated charge information to create a learning dataset. Thus, the learning dataset used in the present embodiment is a plurality of datasets in which solar wind data and charge information are associated with each other every second.

[0169] In the present embodiment, by creating a learning dataset with increased resolution at a resolution of 1 second, the learned model learned using the learning data can further improve the accuracy of charge prediction. Thus, the learning data created in the first embodiment and the learning data created in the second embodiment are different.

[0170] In the present embodiment, using the SSJ data and OMNI data of DMSP-F14 to DMSP-F18 with latitude (MLAT) in the geomagnetic coordinate system of -85 degrees < MLAT < -55 degrees, local time MLT in the geomagnetic coordinate system of 0:00 to 6:00, 18:00 to 24:00 from 1996 to December 31, 2021, a learning dataset is created, and machine learning of a machine learning model is performed using the learning dataset. Also in the present embodiment, the XGBoost is adopted as the machine learning algorithm used for learning.

[0171] Table 6 is a table explaining the SSJ data used in the learning dataset.

[0172]

Table 6

[0173] As shown in Table 6, among the data (samples) obtained by each of the DMSP-F14 satellite, DMSP-F15 satellite, DMSP-F16 satellite, DMSP-F17 satellite, and DMSP-F18 satellite, the number of data (number of surface charging events) that satisfy all of the above conditions (A) to (G), that is, to which the surface charging label of "charged" is assigned, was 5434, 2673, 13661, 8921, and 7641, respectively.

[0174] Also in the second embodiment, during the creation of the learning dataset, the same undersampling as in the first embodiment was performed. That is, the auroral zone was divided by 1 degree in the latitude direction and 1 hour in the local time direction, and among the divided regions, undersampling was performed on the data to which "non-charged" was assigned so that the ratio of the number of data to which "charged" was assigned to the number of data to which "non-charged" was assigned was approximately 1:1.

[0175] [Verification of the trained model]

[0176] Verification was performed on the trained machine learning model (trained model) that was machine-learned using the above learning dataset. In the verification, SSJ data with -85 degrees < MLAT < -55 degrees and MLT of 0:00 to 6:00 and 18:00 to 24:00 from January 1, 2022 to December 31, 2022 was used as inference data. When the performance of the classifier was confirmed by AUC, the AUC score was 0.818, and it was confirmed that a machine learning model with good classification performance could be created.

[0177] Furthermore, it was found that the trained model according to the second embodiment tends to have an increasing AUC score as the surface charging potential increases. Fig. 16 shows the transition of the AUC score for each surface charging potential. The fact that the AUC score increases as the surface charging potential increases means that dangerous surface charging with a high possibility of satellite failure can be detected more accurately. This is quite useful from the perspective of avoiding damage to artificial satellites due to discharge.

[0178] FIG. 17 shows the distribution of AE INDEX for each surface charge potential. In FIG. 17, the surface charge potential is indicated by a thick solid line when 0 V, a thick dotted line when 95 V, a thick dashed line when 139 V, a thick dashed line when 204 V, a thick dashed-dotted line when 300 V, a thin solid line when 440 V, a thin dotted line when 646 V, a thin dashed line when 949 V, a thin dashed-dotted line when 1392 V, and a thin dashed-dotted line when 2040 V. As shown in FIG. 17, as the surface charge potential increases, the distribution of AE INDEX shifts to the right, becoming increasingly different from the distribution of AE INDEX when the surface is not charged. In the trained model of this embodiment, the classification performance improves as the surface charge detection potential increases, which is thought to be related to the fact that the distribution of AE INDEX shifts to the right as the surface charge potential increases. In other words, the trained model of this embodiment is considered to potentially capture the AE INDEX even though it is not used as a feature.

[0179] XGBoost, used to train the machine learning model, can evaluate the importance of features using several indices. FIG. 18 shows feature importance using gain as an index. Gain indicates the average information gain when a feature is used for segmentation. Information gain refers to the reduction in impurity obtained by segmentation. As shown in FIG. 18, Bz and Vx are the top features. It is known that auroras occur when the solar wind's magnetic field is strong in the southward direction or when its velocity toward the Earth is high. The trained model of this embodiment is thought to determine whether an aurora will potentially occur by looking at the features Bz and Vx. Note that when the trained model of the first embodiment was also evaluated using feature importance using gain as an index, evaluation results similar to those of the second embodiment shown in FIG. 18 were obtained. In other words, it is thought that the model of the first embodiment also determines whether an aurora will potentially occur by looking at the features Bz and Vx.

[0180] Therefore, it is considered that the machine learning models of the first and second embodiments internally potentially capture the occurrence of the AE Index and aurora.

[0181] Furthermore, the inventors analyzed the energy spectrogram data of the SSJ data used in the training data and found a correlation between the magnitude of the surface charge potential and the distribution of the charge positions. Specifically, they found that as the surface charge potential increased, the distribution of the charge positions shifted from MLT 18-24 to MLT 0-6. When the surface charge potential was 139 V or less, the charge density increased at the MLT 18-24 position. When the surface charge potential was 204 V and 300 V, the charge density decreased at the MLT 18-24 position and increased at the MLT 0-6 position. When the surface charge potential was 440 V or more, the charge density increased at the MLT 0-6 position. Furthermore, when the surface charge potential was 2040 V or more, the charge density also increased at the MLT 18-24 position. In light of these findings, data from MLTs 0:00-6:00 and 18:00-24:00 was used to create a training dataset used to train the machine learning model according to the second embodiment.

[0182] As described above, in each of the above embodiments, a machine learning model can be constructed using SSJ data and OMNI data that can be obtained from an existing server. Furthermore, using a trained machine learning model, it is possible to accurately predict future satellite charging from OMNI data that can be obtained from an existing server. This eliminates the need to implement charging countermeasures for a space debris capture system, install a plasma generator, install a charging measurement device, etc., and makes it possible to predict satellite charging without placing a burden on the satellite system.

[0183] Furthermore, in each of the above embodiments, in creating a learning data set in which solar wind data and electrostatic charge information are associated with each other for use in machine learning, by specifying the conditions used to determine whether or not an object is electrostatically charged when assigning a surface electrostatic charge label included in the electrostatic charge information, it is possible to construct a machine learning model that can output the possibility of electrostatic charge of an object to be predicted with high accuracy without using aurora information.Then, by using such a machine learning model, it is possible to accurately predict the electrostatic charge of an object to be predicted with high accuracy.

[0184] Third Embodiment

[0185] In the third embodiment, the surface charge labels included in the charge information of the training data set used to construct the learning model are assigned using a method different from that of the first and second embodiments. The third embodiment differs significantly from the first and second embodiments in that the method for acquiring these surface charge labels is different, but the other information in the training data set used to construct the machine learning model is the same. Furthermore, in the third embodiment, the configurations of the learning device and the information processing device using the trained machine learning model are the same as those of the first and second embodiments. Furthermore, the machine learning algorithm used to construct (train) the machine learning model is XGBoost, the same as in the first embodiment.

[0186] In the third embodiment, as in the first and second embodiments, in creating a learning dataset used to build a machine learning model, the solar wind data and the electrification information are associated with each other based on the time (date) at which the solar wind reaches the Earth. In other words, the learning dataset is created by associating solar data and electrification information on the same date.

[0187] The following description will focus on the difference from the first and second embodiments, "assignment of surface charge labels included in the charge information of the training dataset," and will omit a description of the same configuration as the first and second embodiments. In this embodiment, surface charge labels are assigned by an image classification model constructed using an energy-time spectrogram of ion differential number flux included in SSJ data from the DMSP satellite.

[0188] [Assigning surface charge labels included in the charge information of the training dataset]

[0189] In the third embodiment, an image classification model is used to assign a surface charge label included in the charge information. The image classification model classifies an energy-time spectrogram of an ion differential number flux as "charged" or "uncharged." The image classification model can be constructed by training a convolutional neural network (CNN) using a training dataset including a corresponding energy-time spectrogram of the ion differential number flux and surface charge label information (information of "charged" or "uncharged"). The energy-time spectrogram serving as the training image is a visualized image. When charged, a "characteristic spatial structure" appears in the energy-time spectrogram of the ion differential number flux. The appearance of a "characteristic spatial structure" means that the condition (J) described below is satisfied. In other words, in the third embodiment, condition (J) is the charge condition for determining "charged" when assigning a surface charge label. If condition (J) is satisfied, in other words, if a "characteristic spatial structure" appears in the energy-time spectrogram, the surface charge label "charged" is assigned to the energy-time spectrogram of the ion's differential particle flux. On the other hand, if condition (J) is not satisfied, in other words, if a "characteristic spatial structure" does not appear in the energy-time spectrogram, the surface charge label "uncharged" is assigned to the energy-time spectrogram of the ion's differential particle flux.

[0190] (J) In the energy-time spectrogram (specifically, the energy-time spectrogram of the ion differential number flux), the differential particle flux is 10 2 particles / (cm 2 ·sec·str·eV) or more 10 6 particles / (cm 2 sec·str·eV) or less, the differential particle flux is 10 5particles / (cm 2 The condition is that a spatial structure appears that shows a value of σ (σ sec str eV) or more.

[0191] In addition, under condition (J), the differential particle flux is 10 5 particles / (cm 2 The number of energy channels that is greater than or equal to 1 (sec·str·eV) is sufficient.

[0192] The above condition (J) is, in other words, that in the energy-time spectrogram of the differential particle flux of the ion, which is a visualized image, the differential particle flux is 10 2 particles / (cm 2 ·sec·str·eV) or more 10 6 particles / (cm 2 When the color scale is taken in the range of 10 sec str eV or less, the differential particle flux is 5 particles / (cm 2 The condition is that there exists a location with a corresponding color level of at least (·sec·str·eV).

[0193] An energy-time spectrogram is a visualized image in which the color changes depending on the value of the energy flux on the energy axis and the time axis, and has an appearance similar to that of data 9B in Figure 9. The energy-time spectrogram visually shows the temporal change of the energy spectrum. In Figure 9, the higher the energy flux, the denser the dots, and areas with relatively high energy flux are shown in darker colors. Note that data 9B in Figure 9 is the energy-time spectrogram of the ion differential energy flux and the electron differential energy flux, but the energy-time spectrogram of the ion differential number flux also has an appearance similar to that of data 9B. The appearance of the "characteristic spatial structure" in the energy-time spectrogram of the differential particle flux of an ion means a spatial structure that is displayed relatively strongly, for example, as in the energy-time spectrogram of the differential energy flux of an ion shown in data 9B of FIG. 9 (in this embodiment, the energy flux is 10 5 particles / (cm 2 sec str eV). 5 particles / (cm 2 If the state of the voltage (V) or more continues for, for example, several tens of seconds, a dark line structure extending linearly in the horizontal direction appears, as shown in data 9B in FIG.

[0194] Next, the construction of the image classification model will be described.

[0195] In this embodiment, a training dataset used to build an image classification model was created using data from charging events where the above-mentioned "characteristic spatial structure" could be confirmed and data from non-charging events where the "characteristic spatial structure" could not be confirmed, extracted using statistical methods from DMSP-F16 observation data (SSJ data) from 2009 to 2019. Energy-time spectrograms of ion differential particle fluxes with a time axis of 60 seconds were created using the energy-time spectrograms of ion differential particle fluxes contained in the observation data as training images included in the training dataset. In this embodiment, since the SSJ data is data obtained at 1-second intervals and the OMNI data is data obtained at 1-minute intervals, the SSJ data corresponding to each time point in the OMNI data, acquired in 1-minute increments, were used to create energy-time spectrograms of ion differential particle fluxes with a time axis of 60 seconds. In this embodiment, a training data set was created so that when the energy-time spectrogram of the differential particle flux of the ion on the 60-second time axis satisfies the condition (J), a charge information label of "charged" is assigned, and when the condition is not satisfied, a charge information label of "uncharged" is assigned. Note that, although the time axis is 60 seconds in this embodiment, the present invention is not limited to this.

[0196] An image classification model was constructed by training a convolutional neural network (CNN) using the training dataset created in this way, which included the energy-time spectrogram of the ion's differential particle flux and the surface charge label information ('charged' or 'uncharged') associated with the energy-time spectrogram.

[0197] In the first and second embodiments described above, the surface charge label is assigned based on the charging conditions determined by analyzing the plasma environment at the time of charging using a statistical method. However, as in the third embodiment, the surface charge label may be assigned based on an image classification model constructed using an energy-time spectrogram of the differential particle flux of ions.

[0198] The configuration of the third embodiment can further improve the accuracy of charge prediction of the charge prediction target object using a machine learning model compared to the first and second embodiments. That is, for example, in the second embodiment, even if the charging conditions shown in the second embodiment are satisfied (i.e., a surface charge label of "charged"), the "characteristic spatial structure" may not be confirmed on the energy-time spectrogram in some cases, and even if the charging conditions shown in the second embodiment are not satisfied (i.e., a surface charge label of "uncharged"), the "characteristic spatial structure" may be confirmed on the energy-time spectrogram in some cases. In contrast to this, in the third embodiment, the image classification model directly determines whether an object is "charged" or "uncharged" depending on whether a "characteristic spatial structure" that satisfies condition (J) is confirmed on the energy-time spectrogram (visualized image). Therefore, by using a machine learning model trained using a training dataset that includes information on surface charge labels assigned using such an image classification model, it is possible to further improve the accuracy of charge prediction for charge prediction targets.

[0199] In addition to condition (J), if the following condition (K) is also satisfied, the surface charge label "charged" may be assigned to the energy-time spectrogram of the ion differential particle flux, and the surface charge label "uncharged" may be assigned otherwise. Condition (K) means that a dark line structure appears in the energy-time spectrogram. By using this configuration, it is possible to remove data that suddenly appears as abnormal values ​​that are likely to be noise, and to improve the accuracy of charge prediction of an object whose charge is to be predicted using a machine learning model.

[0200] (K) The condition is that the spatial structure of condition (J) appears in the energy-time spectrogram in the same channel for more than 3 consecutive seconds.

[0201] [How to create a training dataset]

[0202] A method for creating a learning data set by the control unit 3 of the learning device 1 according to this embodiment will be described according to the flow of FIG.

[0203] The control unit 3 of the learning device 1 acquires energy-time spectrogram data of the differential particle flux of ions observed by the DMSP satellite (ion energy spectrogram data) (ST31), and creates energy-time spectrogram data of the differential particle flux of ions on a 60-second time axis (ST32).

[0204] Next, the control unit 3 assigns a surface charge label of “charged” or “uncharged” to the energy-time spectrogram of the ion differential particle flux on the 60-second time axis created using the above-mentioned image classification model (ST33), and generates charge information including information on the surface charge label (ST34).

[0205] Next, the control unit 3 acquires the solar wind data 70 shown in Table 2 (ST35).

[0206] Next, the control unit 3 uses the generated electrostatic charge information and the acquired solar wind data to create a learning data set such that electrostatic charge information and solar wind data with the same date are paired (ST36). The control unit 3 stores the created learning data in the learning data set storage unit 41. In this way, a training dataset for a machine learning model can be created.

[0207] Although the present embodiment has been described as an example of constructing an image classification model using an energy-time spectrogram of ion differential particle flux, it is also possible to construct an image classification model using an energy-time spectrogram of electron differential particle flux. However, in the energy-time spectrogram of electron differential particle flux, regions of high energy flux tend not to appear, and characteristic spatial structures tend not to appear. On the other hand, in the energy-time spectrogram of ion differential particle flux, regions of high energy flux tend to appear, and characteristic spatial structures tend to appear. Therefore, it is preferable to construct an image classification model using the energy-time spectrogram of ion differential particle flux data.

[0208] Furthermore, in the third embodiment, an example was given in which an image classification model was constructed using an energy-time spectrogram of ion differential particle flux, but an image classification model may also be constructed using an energy-time spectrogram of ion differential energy flux. Even when a surface charge label is assigned using an image classification model constructed using an energy-time spectrogram of ion differential energy flux data, the accuracy of charge prediction of a charge prediction target object using a machine learning model can be improved.

[0209] In the third embodiment, as in the first and second embodiments, a machine learning model can be constructed using SSJ data and OMNI data that can be obtained from an existing server. Furthermore, using a trained machine learning model, it is possible to accurately predict future satellite charging from OMNI data that can be obtained from an existing server. This eliminates the need to implement charging countermeasures for a space debris capture system, install a plasma generator, or install a charging measurement device, and makes it possible to predict satellite charging without imposing a burden on the satellite system.

[0210] Furthermore, in the third embodiment, in creating a learning dataset in which solar wind data and electrostatic charge information are associated with each other for use in machine learning, an image classification model constructed using CNN is used to assign surface electrostatic charge labels included in the electrostatic charge information. This makes it possible to construct a machine learning model that can output the possibility of electrostatic charge of an electrostatic charge prediction target object with even greater accuracy, even without using auroral information. By using such a machine learning model, it is possible to accurately predict the electrostatic charge of an electrostatic charge prediction target object.

[0211] <Modification> Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and it goes without saying that various modifications can be made within the scope of the gist of the present invention.

[0212] In the above-described embodiment, an example was given in which the information processing device 110 of the information processing system 100 stores the trained model, but the trained model may also be stored in a server separate from the information processing device having the control unit 12.

[0213] Furthermore, for example, the user terminal may be equipped with a control unit 12, or may be an information processing device that executes a series of processes related to presenting predicted electrification information (acquisition of solar wind data, calculation of the possibility of electrification using a trained model, and generation of a display image).

[0214] Furthermore, for example, although only one information processing device 110 is shown in the above embodiment of the information processing system 100, the processing executed by the information processing device 110 may be distributed and executed by multiple information processing devices. For example, the solar wind data acquisition processing and the electrification state prediction processing using the trained model, and the display image generation processing may be executed by separate information processing devices. In this case, the information processing device that executes the display image generation processing may be a user terminal, and if an application program for displaying a display image based on information about the possibility of electrification is installed in the user terminal, the application program may display the display image.

[0215] Among the inventions described in the claims of this application, the inventions described as "learning methods" and "information processing methods" have each step automatically performed by at least one device such as a computer through software-based information processing, and are not performed by a human using a device such as a computer. In other words, the "learning methods" and "information processing methods" are information processing methods using computer software, and are not methods in which a human operates a computing tool called a computer. [Explanation of symbols]

[0216] 1...Learning device 3...Control unit 40...Machine learning model 60...Ion and electron energy spectrogram data 70...Solar wind data (OMNI data) 100...Information Processing Systems 150... Trained model (trained machine learning model)

Claims

1. a data input unit to which a learning data set of solar wind data and electrostatic charge information of an observation satellite, which are associated with each other, is input; a control unit that performs machine learning using the learning dataset to construct a machine learning model that outputs the possibility of charging of an object to be predicted to be charged, based on input solar wind data; and A learning device comprising: The charge information includes information on a surface charge label that indicates whether the surface state of the observation satellite is charged or not, and is assigned based on energy spectrogram data relating to ion flux and / or electron flux observed by the observation satellite. Learning device.

2. the surface charge label is assigned based on the presence or absence of an abnormal value of the electron differential energy flux and / or the ion differential energy flux in a specific energy range of the energy spectrogram data; The learning device according to claim 1 .

3. (A) In the energy spectrogram data, the differential energy flux of the electrons having an energy of 14 keV or more is 10 ８ eV / (cm 2 ・ΔeV・ster・s) or more, and / or (B) The specific energy range is an energy range of 95 eV or more and 2040 eV or less, and the differential energy flux of the ion has an abnormal value. The surface charge label "charged" is assigned to the energy spectrogram data that satisfies the above, and the surface charge label "uncharged" is assigned to the rest. The learning device according to claim 2 .

4. (C) In the energy spectrogram data, an abnormal value of the differential energy flux of the ion is 5×10 ６ eV / (cm 2 ・ΔeV・ster・s) or more The surface charge label "charged" is assigned to the energy spectrogram data that further satisfies the above, and the surface charge label "uncharged" is assigned to the other data. The learning device according to claim 3 .

5. (D) In ​​the energy spectrogram data, there are three or less abnormal values ​​for the differential energy flux of the ion in all energy channels. The surface charge label "charged" is assigned to the energy spectrogram data that satisfies the above, and the surface charge label "uncharged" is assigned to the rest. The learning device according to claim 3 .

6. (D) In ​​the energy spectrogram data, there are three or less abnormal values ​​for the differential energy flux of the ion in all energy channels. The surface charge label "charged" is assigned to the energy spectrogram data that satisfies the above, and the surface charge label "uncharged" is assigned to the rest. The learning device according to claim 4 .

7. (D) In ​​the energy spectrogram data, there are three or less abnormal values ​​for the differential energy flux of the ion in all energy channels. (E) the largest outlier in the energy spectrogram data is a differential energy flux of an ion that is twice or more the second largest outlier; (F) the largest outlier in the energy spectrogram data is in the energy range of 95 eV or more and 2040 eV or less; (G) The largest outlier in the energy spectrogram data is 5 x 10 ６ eV / (cm 2 ・ΔeV・ster・s) or more 10 １０ eV / (cm 2 ・ΔeV・ster・s) or less The surface charge label "charged" is assigned to the energy spectrogram data that satisfies all of the above, and the surface charge label "uncharged" is assigned to the rest. The learning device according to claim 3 .

8. (C) In the energy spectrogram data, an abnormal value of the differential energy flux of the ion is 5×10 ６ eV / (cm 2 ・ΔeV・ster・s) or more The surface charge label "charged" is assigned to the energy spectrogram data that further satisfies the above, and the surface charge label "uncharged" is assigned to the other data. The learning device according to claim 7 .

9. (H) The condition that energy spectrogram data that satisfies all of the above conditions (A) to (G) is observed two or more times consecutively. The surface charge label "charged" is assigned to the energy spectrogram data that further satisfies the above, and the surface charge label "uncharged" is assigned to the rest. The learning device according to claim 8 .

10. the training data set consists only of data when the observation satellite is in the auroral zone; The learning device according to claim 1 or 2.

11. The surface charge label "charged" is assigned to a state in which the surface of the observation satellite is charged, and the surface charge label "uncharged" is assigned to a state in which the surface of the observation satellite is not charged, The learning data set is data in which the distribution of position information of the observation satellite when the charged surface charge label is assigned and the distribution of position information of the observation satellite when the uncharged surface charge label is assigned are matched. The learning device according to claim 10.

12. A learning method for constructing a machine learning model that outputs the likelihood of charging of a target object for charge prediction from input solar wind data by performing machine learning using a learning dataset of solar wind data and electrostatic charge information from an observation satellite that are associated with each other, the method comprising: The charge information includes information on a surface charge label that indicates whether the surface state of the observation satellite is charged or not, and is assigned based on energy spectrogram data relating to ion flux and / or electron flux observed by the observation satellite. How to learn.

13. Learning devices, A step of constructing a machine learning model that outputs the possibility of charging of an object to be predicted from input solar wind data by performing machine learning using a learning dataset of solar wind data and charging information from observation satellites that are associated with each other. A learning program that executes the following: The charge information includes information on a surface charge label that indicates whether the surface state of the observation satellite is charged or not, and is assigned based on energy spectrogram data relating to ion flux and / or electron flux observed by the observation satellite. Learning program.

14. Acquire solar wind data, Calculating the possibility of charging an object to be predicted to be charged from the solar wind data using the machine learning model that has been trained and constructed by the learning device according to claim 1 or 2 Control unit An information processing system comprising:

15. acquiring solar wind data; a step of calculating the possibility of charging of the object to be predicted to be charged from the solar wind data using the machine learning model that has been trained and constructed by the learning device according to claim 1 or 2; An information processing method executed by an information processing device.

16. In the information processing device, acquiring solar wind data; a step of calculating a possibility of charging of an object to be predicted to be charged from the solar wind data using the machine learning model that has been trained and that is constructed by the learning device according to claim 1 or 2; A program that executes the following.

17. A non-transitory computer-readable recording medium storing the program according to claim 16.

18. the surface charge label is assigned based on an energy-time spectrogram of a differential particle flux of the ion, which is energy spectrogram data relating to the flux of the ion; The learning device according to claim 1 .

19. The surface charge labels are assigned by an image classification model constructed by training a convolutional neural network (CNN) using the energy-time spectrogram of the differential particle flux of the ions and the surface charge label information that are associated with each other; In the surface charge label information used when constructing the image classification model, (J) In the energy-time spectrogram, the differential particle flux is 10 ２ particles / (cm 2 ・sec・str・eV) or more 10 ６ particles / (cm 2 sec str eV) or less, the differential particle flux is 10 ５ particles / (cm 2 The condition is that a spatial structure with a value of sec, str, or eV appears. The surface charge label "charged" is assigned to an energy-time spectrogram that satisfies the condition (J), and the surface charge label "uncharged" is assigned to an energy-time spectrogram that does not satisfy the condition (J). The learning device according to claim 18.

20. (K) The condition that the spatial structure of (J) appears in the same channel in the energy-time spectrogram for 3 seconds or more consecutively. The surface charge label "charged" is assigned to an energy-time spectrogram that further satisfies the condition (K), and the surface charge label "uncharged" is assigned to an energy-time spectrogram that does not satisfy the condition (K). The learning device according to claim 19.

Images