In this paper, we introduce automatic control methods that exploit both the accurate prediction performance of mathematical models and the fast calculation of data-driven models.
This study investigates anomaly detection using a model-based approach, specifically applied to JAXA Low-speed Wind Tunnel (LWT1). A model-based numerical model of the LWT1 was developed, and Uncertainty Quantification (UQ) was conducted to estimate the probability distribution of the objective variable under normal conditions. The 95% confidence interval of the distribution was defined as the normal space for anomaly detection. A demonstration test campaign introducing synthetic anomaly was conducted. The experimental results demonstrate that the model-based approach enables effective anomaly detection even in systems with limited training data for machine-learning.
In spacecraft thermal design, constructing an accurate thermal mathematical model is crucial for reliably predicting the temperatures of various components. However, current practices involve manually adjusting model parameters based on thermal vacuum test results, which is both time-consuming and costly. This study proposes the use of simulation-based inference (SBI) for estimating the parameters. SBI offers the flexibility to handle complex and nonlinear models. Additionally, it provides posterior distributions for the estimated parameters, enabling a quantitative assessment of uncertainties and confidence intervals. To validate the effectiveness of this approach, numerical experiments are conducted using a small satellite model under simplified conditions.
工場やビルの設備には組込み保全運転という例外制御が装備されている場合があり,リアルタイム電力料金適応制御など将来のエネルギー集中管理システムの観点からは突発的な外乱となる。本研究では,突発的な組込み保全運転の発生を5分前に予測する時系列モデルの機械学習を検討した。一例として数時間の時系列履歴に依存する場合の実測データから予測モデルを Long Short Term Memory (LSTM) ニューラルネットワークとして実装した。シミュレーション実験の一例として,モデル予測精度0.93およびスレットスコア0.6となり,りアルタイム電力料金管理への効果は27%改善が期待できる結果を得た。
The health management of artificial satellites is becoming increasingly important for enhancing operational efficiency and enabling early change detection. In this study, we systematically define and classify the complex and unique characteristics of satellite telemetry data. Furthermore, we introduce Quantile Gaussian Process Regression (Q-GPR) as an evaluation method capable of addressing these challenges. Utilizing real satellite data, we demonstrate that Q-GPR is effective in anomaly detection even for data with coarse quantization levels and datasets that involve multiple operational modes, which have been challenging for conventional methods to handle.