[4L2-GS-13] AI application: Anomaly detection and maintenance

[4L2-GS-13-05] Rocket Engine Fault Detection and Diagnosis Using Monte Carlo Simulation

〇Noriyasu Omata¹, Daiwa Satoh¹, Miki Hirabayashi¹, Seiji Tsutsumi¹, Kaname Kawatsu¹, Masaharu Abe² (1. Japan Aerospace Exploration Agency, 2. Ryoyu Systems Co., Ltd.)

Keywords:Monte Carlo simulations, model based methods, fault detection, fault diagnosis, sensor optimization

In fault detection of complex systems, there exists a model-based approach using simulation techniques in addition to a data-driven approach. Since the simulation results are unique in general, model-based approaches cannot consider the effect of imperfect modeling and real system uncertainties. In this study, fault detection which avoids this problem by combining Monte Carlo simulation with data-driven techniques is carried out. The subject is fuel leakage in a rocket engine, which is difficult to detect but is a serious failure mode. A data set which consists of multiple patterns of both normal operation and fuel leakage were generated by Monte Carlo simulation. Although the faults in this data set cannot be detected by conventional methods, they can be detected to some extent by multivariate machine learning methods such as SVM. In addition, this paper shows the method of the sensor selection by the greedy method and an attempt for the diagnosis of the leakage position.