Deep learning for turbojet engine prognostics: evaluating RNN and transformer-based models for EGT forecasting

Time series forecasting is vital in aviation for optimizing engine efficiency and decreasing operational expenses. This work investigates the predictive effectiveness of deep learning models for forecasting Exhaust Gas Temperature (EGT) in the Viper 632-43 turbojet engine, an essential parameter in engine health monitoring and sustainability. While NARX models have been widely employed, their limits in capturing long-term dependencies demand new techniques. We examine RNN-based architectures (LSTM, Bidirectional LSTM, Stacked LSTM, CNN-LSTM, and GRU) and transformer-based (Informer, Autoformer, PatchTST, TimesNet, TFT) models. Results reveal that LSTM achieved the lowest RMSE (35.05) and MAE (17.15), while TimesNet achieved an RMSE of 35.48 and R2 of 0.9076, outperforming the baseline NARX model. These findings stress the relevance of model selection based on forecasting needs, enhanced predictive maintenance strategies, and fuel efficiency benefits in aviation, directing future applications in engine health monitoring.