An evaluation of the Long Short-Term Memory model for predictive maintenance applications in the aircraft industry

This study investigates the prediction of remaining useful life (RUL) in the aircraft industry using Long short-term memory (LSTM). With LSTM, there are challenges in optimising the choice of structure, type of architecture, number of neurons, number of hidden layers, and learning parameters. Causal-comparative research investigates the impact of hyper-parameter changes in LSTM to predict RUL. NASA C-MAPSS FD001 dataset was processed, and LSTM models use different hyper-parameters for window size, number of units, and dropout rate. The LSTM models predict the RUL and a Root mean square error (RMSE). RMSE comparisons for Linear regression (LR), Random Forest (RF), and Decision tree (DT) models trained using the Time series feature extraction library (TSFEL) were made. Increasing window size, or number of units, reduces the RMSE. A lower dropout rate resulted in lower RMSE. The best performance was an RMSE of 14,34. Using TSFEL resulted in a 2% improvement in LR, a 13% in DT and a 1% in RF RMSE; however, not better than the performance of LSTM. Study shows that using TSFEL for feature extraction improves the performance of traditional models.