11:00 AM - 1:00 PM
[PEM12-P15] Control of switching two observation modes on FM-CW ionospheric observation system based on reinforcement learning
Keywords:FM-CW radar, reinforcement learning, LSTM (long-short term memory), supervised machine learning
Equatorial Plasma bubbles (EPBs) of the ionospheric irregularities are known to cause GPS positioning errors and radio wave propagation abnormalities. Recently, numerical experiments have advanced the understanding of the local generation mechanism of EPBs, but the development of environmental field controlling the generation and suppression of EPBs has not yet been fully clarified. This study aims to reveal observationally the structure and generation mechanism of environmental fields in inner-magnetosphere and ionosphere that is linked to the development of EPBs.
We focus on a three-dimensional coupling system of ionospheric E-F regions controlling equatorial jet current (EEJ) as a model that connects the pre-sunset EEJ, pre-reversal enhancement (PRE) at near sunset, and EPBs after sunset. In order to detect this coupling system, we have developed a multi-ionospheric observation project with FM-CW (Frequency Modulated Continuous Wave) radar, MAGDAS (MAGnetic Data Acquisition System) magnetometer network and SDR-based scintillation detector. The FM-CW radar has two kinds of observation modes: one is Ionosonde mode, and the other is Doppler mode. FM-CW radar enables continuous multi-mode ionospheric observation by switching between the detection of time evolution from PRE to plasma bubble by Ionosonde mode and the observation of F region electric field by Doppler mode. We have developed a new "autonomous FM-CW control system" without the manual operation schedule. The new FM-CW system consists of the supervised machine learning and reinforcement learning by using several ionospheric disturbance triggers.
First, we create a pre-trained discriminant model using LSTM (long-short term memory), which is one of the supervised learning methods. This model takes as input the time-series observation data of the ionosphere, solar wind, and magnetosphere, which may be involved in the generation of plasma bubbles and classifies them into (a) ionospheric steady state (b) plasma bubble generation period. In addition, the pre-training discriminant model is incorporated into the reinforcement learning, which is a framework for controlling the mode switching of the FM-CW observation system. Reinforcement learning learns a policy to obtain the most reward through a series of actions. In the Normal mode (a), the ionospheric density as a function of altitude is measured by the ionosonde mode. When EPBs occurs (b), the ionospheric electric field is estimated by measuring the vertical motion of the plasma with the Doppler mode. As a result of the experiment, the accuracy of the pre-training discriminant model was 62.6% for classifying (a) ionospheric steady state and (b) plasma bubble generation period. When reinforcement learning was applied to the pre-training model, the classification accuracy was 53.2%. The low accuracy may be due to the fact that the system did not learn enough to select the appropriate observation mode in the reinforcement learning process. In the presentation, we will explain the new FM-CW control system and show the result of examination.
We focus on a three-dimensional coupling system of ionospheric E-F regions controlling equatorial jet current (EEJ) as a model that connects the pre-sunset EEJ, pre-reversal enhancement (PRE) at near sunset, and EPBs after sunset. In order to detect this coupling system, we have developed a multi-ionospheric observation project with FM-CW (Frequency Modulated Continuous Wave) radar, MAGDAS (MAGnetic Data Acquisition System) magnetometer network and SDR-based scintillation detector. The FM-CW radar has two kinds of observation modes: one is Ionosonde mode, and the other is Doppler mode. FM-CW radar enables continuous multi-mode ionospheric observation by switching between the detection of time evolution from PRE to plasma bubble by Ionosonde mode and the observation of F region electric field by Doppler mode. We have developed a new "autonomous FM-CW control system" without the manual operation schedule. The new FM-CW system consists of the supervised machine learning and reinforcement learning by using several ionospheric disturbance triggers.
First, we create a pre-trained discriminant model using LSTM (long-short term memory), which is one of the supervised learning methods. This model takes as input the time-series observation data of the ionosphere, solar wind, and magnetosphere, which may be involved in the generation of plasma bubbles and classifies them into (a) ionospheric steady state (b) plasma bubble generation period. In addition, the pre-training discriminant model is incorporated into the reinforcement learning, which is a framework for controlling the mode switching of the FM-CW observation system. Reinforcement learning learns a policy to obtain the most reward through a series of actions. In the Normal mode (a), the ionospheric density as a function of altitude is measured by the ionosonde mode. When EPBs occurs (b), the ionospheric electric field is estimated by measuring the vertical motion of the plasma with the Doppler mode. As a result of the experiment, the accuracy of the pre-training discriminant model was 62.6% for classifying (a) ionospheric steady state and (b) plasma bubble generation period. When reinforcement learning was applied to the pre-training model, the classification accuracy was 53.2%. The low accuracy may be due to the fact that the system did not learn enough to select the appropriate observation mode in the reinforcement learning process. In the presentation, we will explain the new FM-CW control system and show the result of examination.