A Temporal Self-Attention Guided Deep Learning Approach for Prognosis of Remaining Useful Life of Aircraft Turbofan Engines

In the aviation industry, a reliable prognosis of the remaining useful life (RUL) of aircraft engines is a prerequisite for effective predictive maintenance as well as averting critical failures. But the high non-linearity of engine components and operation in complex environments make it more challenging to capture the deterioration behavior and estimate the RUL. To deal with this concern, a deep learning ensemble model with a Temporal Self-Attention Mechanism (TSAM) is presented to estimate the RUL of aircraft turbofan engines effectively. Firstly, the importance of different sensors is evaluated by applying the TSAM method to the raw sensor data. Then Temporal Convolutional Network (TCN) is applied to the weighted sensor data to extract high-dimensional features. Next, a Bi-directional LSTM network is used to acquire the long-term dependencies in the time series. To further improve the performance we used the Bayesian Optimization strategy to optimize the best hyperparameters. Finally, the efficiency of the proposed TSAM-TCN-Bi-LSTM model is verified using the Commercial Modular Aero Propulsion System Simulation (C-MAPSS) dataset obtained by NASA and evaluated against other existing methods. The study demonstrates that the presented ensemble model can prominently yield better results than other approaches for estimating the remaining useful life of aircraft turbofan engines.