Fault Analysis Of Turbopump Test Data Using Novelty Detection Technology

Lacking of fault samples is a knotty problem in the fault detection of complex systems such as Liquid-propellant Rocket Engines (LRE). Generally, if novel cases are very rare or there are no fault samples, it's difficult to analyze the fault patterns. In this case, health monitoring and fault diagnosis should turn to Novelty Detection(ND) technology, which offers a solution to this problem by modeling normal data, using some distance measure and a threshold for determining novelty.This thesis describes the exploratory work of applying a powerful ND technology– One-Class Support Vector Machines (OC-SVMs) to the detection of faults embedded in turbopump vibration data. The main contents of this thesis include:(1) Many popular ND methods are introduced together with their characters and two kinds of OC-SVMs methods are described in detail. Based on these descriptions, a ND model called Support Vector Data Description (SVDD) is founded.(2) A qualitative guide for setting those parameters in OC-SVMs is investigated. A multi-layer high-speed training strategy was proposed to enable support vector algorithm to handle large training data.(3) Features of turbopump test data both in time domain and in frequency domain are analyzed using statistical estimation methods. The value of applying these features into ND is discussed.(4) Parameters in the OC-SVMs model are modified according to extracted detection features and many turbopump test data are detected using this model.(5) The performance of ND model is also validated through the vibration signals simulated in laboratory rotor system.With the analysis of both turbopump test data and laboratory rotor system vibration signals, the results show that ND technology is able to detect vibration faults successfully.