Critical modeling issues for prediction of turbine performance degradation - Use of a stochastic-neuro-fuzzy inference system

Abstract

One key aspect when developing a robust health management system for turbines is the development of accurate and robust fault classifiers. The paper illustrates the application of a hybrid Stochastic-NeuroFuzzy-Inference System to fault diagnostics and prognostics for turbine performance. The random fluctuations of turbine performance parameters in different varying operating conditions are modeled using a multivariate stochastic model. The fault diagnostic and prognostic are computed using a stochastic flow-path analysis model of the turbine. At any time, the fault risk condition is approached as a conditional reliability problem based on the measurement of parameter deviations from the normal operating condition. The paper illustrates the application of the proposed system to a typical aircraft turbofan engine for in-flight engine performance diagnostic and prognostic. 1.0Turbine Performance Modeling Figure 1 shows a sketch of a typical turbofan engine including the performance parameters considered herein for fault diagnostic and prognostic. Figures 2 and 3 show pressure variations as a function of the high-pressure shaft speed stationary conditions versus highly transient operating conditions, respectively. It is obvious from these figures that although for slow varying conditions the pressure closely follows a nonlinear relationship with shaft speed, for highly transient operating conditions the pressure deviates from this nonlinear path due to highly transient conditions and significant changes in the inlet conditions, namely inlet pressure, temperature and mass flow. This means using deviations from a fitted polynomial regression line for diagnostics, as commonly used in engine health monitoring application based on ground-test data, is not suited to in-flight conditions. In fact the large stochastic variability projected on the pressure-speed plane in Figure 3 is only apparent. This large variability is mostly due the transient variations induced by the pilot maneuvers. A key aspect for getting realistic predictions for in-flight operating conditions is to separate the true statistical variabilities (random part) from the functional variabilities introduced by engine transient behavior. For fast transient conditions the functional dependence between turbine performance parameters becomes complex and highly nonlinear. If these transient functional dependencies between multiple parameters are ignored then the statistical variability is overestimated and the computed fault risks are unreliable, being highly overestimated as shown in Figure 3.As physics-based, analytical GPA models only cater for quasi-stationary engine operation, an alternative scheme capable of including the highly transient in-flight conditions has been developed. This lead to the formation of a stochastic diagnostic-prognostic model based on parameter statistical deviations from an adaptive networkbased fuzzy inference system GPA model. This model was developed by calibrating the analytical

Performance-analysis-based gas turbine diagnostics: A review

“…A static pattern analysis approach was proposed by Patel et al [79] and Arkov et al [90], where the observation of gas turbine status was expressed by a probability density or histogram approach and any deviation of the engine from its normal condition can be indicated by a low likelihood of the observation. A probabilistic fault diagnostic approach was introduced by Ghiocel and Roemer [124] and Roemer and Ghiocel [125] and was further described by Ghiocel and Altmann [126] and used by Roemer et al [127]. In the method, both the monitored and fault data uncertainties were considered and described with PDFs.…”