Predicting Jet Engine Failure: A Deep Learning Approach to Remaining Useful Life Estimation

Abstract

Unplanned equipment failure in safety-critical fields such as commercial aviation can carry consequences beyond repair costs: cascading flight cancellations, grounding aircraft, and in severe cases, loss of life. These risks motivate the development of accurate predictive maintenance algorithms that estimate Remaining Useful Life (RUL) from sensor data to intervene before failure occurs. This thesis investigates which Long Short-Term Memory (LSTM)-based recurrent architectures are most effective for turbofan jet engine RUL prediction, using NASA's C-MAPSS dataset of 100 simulated trajectories. Four architectures are implemented and evaluated, each motivated by a specific structural limitation of the preceding model. The central novel contribution is a Ridge regression stacking ensemble that combines the two strongest individual models by exploiting their complementary error profiles. The BiLSTM and Attention-LSTM achieve a mean average error (MAE) of 12.08 and 11.70 flight cycles respectively, competitive with published benchmarks. A binned analysis shows the BiLSTM is stronger near failure, while the Attention-LSTM is stronger at longer horizons where degradation signals are weaker. The stacking ensemble outperforms both models on every metric, with a MAE of 10.51 cycles and a NASA score of 3.55 - a 48% reduction in operational risk over the best single model. The central conclusion is that no single LSTM architecture dominates across RUL ranges, and model selection should be guided by which part of the RUL distribution is most operationally critical. The stacking ensemble demonstrates that combining architectures with complementary error profiles produces marked improvement over any individual model, and represents a promising direction for future scholarship in data-driven prognostics.