Transient and steady state data may contain the same essential fault information but some faults have been shown to be more easily detectable from transient data because the transient records provide significant diagnostic content especially as the fault effects are magnified under transient. Various traditional and conventional techniques such as fault trees, fault matrixes, gas path analysis and its variants have been applied to gas path fault diagnosis of gas turbines. Recently, artificial intelligence techniques such as artificial neural networks (ANN) as well as optimization techniques such as genetic algorithm (GA) are being explored for fault diagnosis activities. In this paper, a novel approach to gas path fault diagnosis is proposed. The method involves the use of ANN with engine transient data. A set of nested neural networks designed to estimate independent parameter (efficiencies and flow capacities) changes due to faults within single or multiple components of a turbofan engine are presented. The approach involves classification and approximation type networks. Measurements from the engine are first assessed by a trained network and if a fault is diagnosed, are then classified into two groups — those originating from sensor faults and those from component faults, by another trained network. Other trained networks continue the fault isolation process and finally the magnitude of the fault(s) is quantified. A computer simulation of the process shows that results from a batched process of these networks can be obtained in less than three seconds. Four of the gas path components — intermediate pressure compressor (IPC), high pressure compressor (HPC), high pressure turbine (HPT) and low pressure turbine (LPT) — and measurements from eight sensors are considered. Sensor noise and bias are also considered in this analysis. The comparison of fault signatures from a steady state and transient process show that diagnosis with transient data can improve the accuracy of gas turbine fault diagnosis.
Asymptotic Prediction Error Variance for Feedforward Neural Networks
