An Adaptive-Tunable-Based Hybrid RBF Network for EGTM Prediction

Yishou Wang,Xianping Zeng,Yanyun Tian,Hanlin Sheng,Yuan Liu

doi:10.1109/access.2021.3052465

Abstract

Aero-engine exhaust gas temperature margin (EGTM) is one of the main indexes of engine replacement; however, the application of existing methods in EGTM forecasting is restricted because of the limited prediction accuracy and many non-linearities. In this study, an adaptive-tunable-based hybrid radial basis function (RBF) network is proposed to improve the prediction accuracy of aero-engine EGTM. Firstly, a hybrid RBF network consisting of a RBF network and a linear regression model is built as a fundamental EGTM predictive algorithm. Secondly, to increase the network’s adaptation capabilities, the structural parameters of the proposed network are adaptively modulated by Brownian motion modeling and particle filter without physics-based models. Finally, multiple sets of EGTM data from a certain type aero-engines in an airline company is selected for engine removal time prediction. Experiment results demonstrate that the proposed adaptive-tunable-based hybrid RBF network with a high prediction accuracy, and can reflect the characteristics of EGTM well and truly, which can capture the dynamic nature of EGTM in time during the forecasting process.

Highlights

The predictive replacement of aero-engine is an important means to improve the economics and safety of the entire fleet
A hybrid radial basis function (RBF) network consisting of a RBF network and a linear regression model is built as a fundamental exhaust gas temperature margin (EGTM) predictive algorithm
The results demonstrated that the proposed model possess a minor mean square error (MSE), MAE and maximum error (MAX) error compared with the fixed-structure RBF

Summary

INTRODUCTION

The predictive replacement of aero-engine is an important means to improve the economics and safety of the entire fleet. Because of the problem that the neural network structure is too large or too small, this paper uses particle filtering to dynamically adjust the hybrid RBF network, that is, as new samples are continuously added, the network structure parameters Nt , Rt , cs,t , σs,t , βt are dynamically adjusted online to capture the dynamics nature’s change of aero-engine EGTM in time. The state transfer equation of the online structure variable RBF network prediction model is as follows: Nt. σs,t = σs,t−1 + εσ , s = 1, 2, · · · , Nt βt = βt−1 + εβ wherein, round(·) means rounding, and assuming εσ ∼ N (0, δσ INt ×Nt ), εc ∼ N (0, δc2I(Nt ∗m)×(Nt ∗m)), εβ ∼ N (0, δβ I(Nt +Rt +1)×(Nt +Rt +1)), where m represents the featural dimension of the input, and δc, δσ2 , δβ is all constant, I(Nt ∗m)×(Nt ∗m), INt ×Nt , I(Nt +Rt +1)×(Nt +Rt +1) is the unity matrix. Due to the different number of cycles before each engine failure, a total of 30 groups of EGTM data sets with different sample numbers were obtained, with totalling1564 EGTM data

PREDICTIVE MODEL EVALUATION CRITERIA

EXPERIMENT AND RESULT ANALYSIS

Findings

CONCLUSION