Fault Diagnosis for Spacecraft Using Probabilistic Reasoning and Statistical Learning with Dynamic Bayesian Networks
- Paper number
IAC-05-D5.2.04
- Author
Mr. Yoshinobu Kawahara, University of Tokyo, Japan
- Coauthor
Dr. Kazuo Machida, University of Tokyo, Japan
- Coauthor
Dr. Takehisa Yairi, University of Tokyo, Japan
- Year
2005
- Abstract
Fault diagnosis is a critical task for the operation of spacecraft. According as spacecraft become more complex and large-scale along with the diversification of needs, the operator's prompt and accurate decisions are indispensable in unusual situations. So the spacecraft diagnosis system should exhaustively detect anomalies, present information for operators to estimate its part and cause, and perform these tasks adaptively to the actual behavior of spacecraft. In other word, the method to "widely, deeply and adaptively" diagnosis is required. A conventional approach to the diagnosis problem for spacecraft is devided into two types, one based on prior-knowledge and data-acquire knowledge. The first type of approach contains the expert system and the model-based diagnosis, and enables detail cause investigation. However, the range of diagnosis is limited to the given diagnosis knowledge and the inaccuracy of the knowledge directly leads the reliability decrease in the diagnosis process. In second type of approach, the diagnosis is performed by constructing some behavioral models with the observation data acquired in the past and monitering newly acquired data using this model. This method enables anomaly detection that reflects spacecraft's actual behavior, however, because the part or cause of detected event cannnot be specified, it is difficult to present the reason of the result to operators. In otehr words, the former enables "narrow and deep" diagnosis by using diagnosis models deductively and the latter "wide, adaptive and shallow" by aqcuiring one inductively. However, these two polar approaches have the difficulty for covering the other' property. We propose "a diagnosis method using probabilistic reasoning and statistical learning" for the purpose of the spacecraft diagnosis system which adaptively enables a wide range of anomaly detection and advanced cause investigation. In our approach, Dynamic Bayesian Networks (DBNs), a general state space model, is constructed as the diagnosis model with prior-knowledge such as dynamics and design information. This DBN is corrected or partly acquired by statistical learning with observation data with EM algorithm. Therefore, the "width" of diagnosis and the "adaptability" are achieved. And an actual "deep" diagnosis which contains anomaly detection and cause investigation is performed by probabilistic reasoning with DBN. This inference task is performed using Rao-blackwellised Particle Filters (RBPFs). The proposed method was applied to the telemetry data (presented from Japan Aerospace Exploration Agency: JAXA) that simulates anomalies of thrusters in rendezvous maneuver of spacecraft. The goal of this experiment is to detect thrust's decrease and present information for operators to estimate which thruster have a breakdown and how behaves in the case that it is difficult for operators to judge the situation. In this experiment, we constructed the DBN with the rough prior knowledge, linearlized Euler's equation and Hill's equation. However, the results was very accurate by means of statistical learning and the effectiveness of our method was confirmed. First, the proposed method can calculate the probabilities that each thruster behave unusual, and the results show that our method can narrow the thrusters broken down down to a few candidates in real-time. Then in our approach, it is possible to estimate the transition of states which are not directly observed such as thrusts of each thruster, and the results of this calculation support for operators to judge the situations.
- Abstract document
- Manuscript document
IAC-05-D5.2.04.pdf (🔒 authorized access only).
To get the manuscript, please contact IAF Secretariat.