Remaining useful life predictions of aero-engines using dilated LSTM with attention mechanism

Prognostics and Health Management (PHM) is a pivotal field that focuses on predicting the future health and operational life of systems and components. One of the key aspects of PHM in aviation is predicting the Remaining Useful Life (RUL) of aircraft engines. Accurate RUL predictions enable airlines and maintenance crews to schedule timely interventions, minimizing the risk of unexpected failures and optimizing maintenance schedules. This paper proposes a novel integration of Dilated Long Short Term Memory with an Attention mechanism (DLSTM+A) for RUL prediction of aircraft engines, addressing the limitations of existing approaches in capturing long-term temporal dependencies. Unlike traditional LSTM models that struggle with long-range dependencies and conventional Dilated RNNs that suffer from gradient issues with simple RNN cells, our approach combines the multi-scale temporal modeling capabilities of dilated connections with the superior memory retention of LSTM cells and the selective focus of attention mechanisms. The methodology consists of two primary phases: data preparation with distribution-driven feature selection and model development with hyperparameter optimization. Using the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset, our DLSTM+A model demonstrates significant improvements over state-of-the-art methods: 56% reduction in RMSE and 43% reduction in score function compared to standard LSTM, and 46% improvement in RMSE and 30% improvement in score function compared to LSTM with Attention. The model achieves the lowest RMSE (7.54) and score function (223) among all evaluated approaches, establishing a new benchmark for RUL prediction accuracy in turbofan engines while maintaining computational efficiency through dilated skip connections.