Evaluation of the impact of Long Short-Term Memory parameters in RUL prediction for aero engines

The paper deals with the application of a Long Short-Term Memory (LSTM) to estimate the Remaining Useful Life (RUL) of NASA engines. The engine's data were retrieved from well-known, free available, datasets, in which 21 variables are monitored during the entire engine life, until failure. Thus, the LSTM was identified as the best method to estimate the engine's RUL because it is based on learning from the data series. The LSTM was developed in Python programming language, using a training dataset for the learning phase and a test dataset to evaluate the performance of the algorithm, both provided by NASA. Then, a sensitivity analysis was carried out to evaluate the impact of three parameters on the Mean Squared Error (MSE) of the RUL and the training computational time, namely: i) the window size, i.e. the number of observations to consider to make predictions; ii) the batch size, i.e the number of samples considered for updating the internal model parameters; iii) the Pearson coefficient, used in the pre-processing phase to identify the most useful variables to give as input to the LSTM algorithm. The results highlighted that the window dimension is the most influential parameter among those considered.