A step parameters prediction model based on transfer process neural network for exhaust gas temperature estimation after washing aero-engines

Abstract

The prediction of Exhaust Gas Temperature Margin (EGTM) after washing aero-engines can provide a theoretical basis for airlines not only to evaluate the energy-saving effect and emission reduction, but also to formulate reasonable maintenance plans. However, the EGTM encounters step changes after washing aeroengines, while, in the traditional models, a persistence tendency exists between the prediction results and the previous data, resulting in low accuracy in prediction. In order to solve the problem, this paper develops a step parameters prediction model based on Transfer Process Neural Networks (TPNN). Especially, “step parameters” represent the parameters that can reflect EGTM step changes. They are analyzed in this study, and thus the model concentrates on the prediction of step changes rather than the extension of data trends. Transfer learning is used to handle the problem that few cleaning records result in few step changes for model learning. In comparison with Long Short-Term Memory (LSTM) and Kernel Extreme Learning Machine (KELM) models, the effectiveness of the proposed method is verified on CFM56-5B engine data.