Parameter modelling of fleet gas turbine engines using gated recurrent neural networks

Abstract

Oriented by the gas-path fault diagnosis of aircraft engines, this work presents a novel parameter modelling scheme for fleet gas turbine engines based on gated recurrent neural networks. Four dynamic models are constructed based on the long short-term memory (LSTM) network and gated recurrent unit (GRU) network to predict the total temperature at the high-pressure turbine outlet of gas turbine engines. We apply the dynamic networks to model the engine parameters by solving a sequence-to-feature regression problem. The proposed scheme is assessed through a comprehensive comparison study on training performance, validation performance, generalization, robustness against noise, and extrapolation performance. The results showed that the proposed network models performed a promising performance on parameter modelling. For the proposed four network models, the NARX-based models have better one-step-ahead prediction performance, while the state-space-based models have better long-term prediction performance. The discussion of the results also showed that the GRU network was an effective data-driven model for parameter modelling of gas turbine engines.