Application of CNN-IndRNN model combined with residual structure in aircraft engine EGT prediction

Abstract

The convolutional neural network (CNN), independently recurrent neural network (IndRNN), and residual structure are used to establish a deep learning model for predicting the exhaust gas temperature (EGT) of aircraft engines. Compared to traditional RNN, the IndRNN has an independent recurrent weight matrix in each unit unfolded in the time dimension, which can more efficiently capture spatiotemporal information. Meanwhile, CNN can capture salient features from the multi-feature and multi-time-step data of the original input. Subsequently, the salient features extracted are combined with the initial input using the residual structure before being fed into the IndRNN module. A case study is conducted on the flight data of two civil aircraft engines to verify the effectiveness of IndRNN compared to traditional RNN. The proposed CNN-IndRNN model can accurately predict the EGT of aircraft engines, with a mean absolute relative error below 1.0%.