[2I4-GS-2] Machine learning: Anomaly detection

[2I4-GS-2-05] Technique for Generating Training Data by Fusing Experiment and Simulation for Fault Detection and Diagnosis of Rocket Engine

〇Daiwa SATOH¹, Miki HIRABAYASHI¹, Seiji TSUTSUMI¹, Noriyasu OMATA¹, Masaharu ABE², Kaname KAWATSU¹ (1. Japan Aerospace Exploration Agency, 2. Ryoyu Systems Co., Ltd.)

Keywords:Model-based methods, Fault detection, Fault diagnosis, System-level simulation

To generate training data for fault detection and diagnostics with machine learning, a System-Level Simulation (SLS) that enables to simulate the global behavior of the reusable rocket engine has been developed. However, the SLS cannot simulate any sensor behaviors due to complicated physical phenomena. To obtain such sensor behaviors with low cost, the sensors that the SLS can simulate are used as explanatory variables for a regression model, and the sensors that the SLS cannot simulate are predicted by the trained model. Meanwhile, the time series of the response variable are nonlinear behavior, so that a single linear regression model no longer predicts it. However, training a nonlinear regression model, such as SVR and Neural Network, are time consuming. In the present study, Ridge regression is used as the regression model, and the training data for the regression model are split into more than one cluster by a Gaussian mixture model before predicting the response variable. By training the Ridge regression models on each cluster, multi regression models predict the nonlinear sensor behavior.