An Investigation of Artificial Neural Network (ANN) in Quantitative Fault Diagnosis for Turbofan Engine

Abstract

This paper is aimed at investigating two kinds of Artificial Neural Network (ANN) applied to quantitative fault diagnosis of turbofan engine gas path components. Among them, one is Back Propagation neural Network (BPN) and the other is Adaptive Probabilistic Neural Network (APNN). Using BPN in order to achieve quantitative fault diagnosis, number of training samples will increase greatly which may lead to the difficulty of iteration convergence. A new learning rule named hybrid rule is introduced to avoid the algorithm falling into static areas and expedite convergence. Recently, a new method to improve the adaptability of multi-layer feed-forward neural network has been developed by the application of Radial Basis Function (RBF). In this paper, the APNN is put forward based on the theory of radial basis function, Bayesian estimation and normal distribution hypothesis of information. It is proposed that the adaptability of APNN can be obtained by applying maximum-likelihood estimation of the output of test case based on a posteriori probability of its input. The investigation shows that BPN and APNN have their own advantages and disadvantages. BPN has faster diagnostic speed and fits the requirement of quantitative diagnosis for single fault. APNN is more adaptive and fit better to quantitative diagnosis for multiple faults.