Abstract
This study evaluates Long short-term memory (LSTM) for predictive maintenance in the aircraft industry by predicting the remaining useful life (RUL) using historical time-series machine sensor data. When utilising deep learning models such as LSTM, there are challenges in determining the optimal choice of structure, type of architecture, the optimal number of neurons, the number of hidden layers, and the optimization of the learning parameters. Causal-comparative research approach is used to investigate impact of hyper-parameter changes in LSTM to predict RUL.
The NASA C-MAPSS predictive maintenance dataset was processed, and LSTM models were trained using different hyper-parameters for the window size, the number of units, and the dropout rate. The number of layers of the LSTM model is kept constant. The LSTM models are developed to predict the RUL then, Root mean square error (RMSE) and R-Squared (R2) values are determined. Comparisons of the RMSE and R2 were made with LSTM, Linear regression (LR), Random forest (RF), and Decision tree (DT) models trained with features extracted using the Time series feature extraction library (TSFEL).
Results indicated that increasing the window size, or number of units reduces the RMSE. A lower dropout rate resulted in lower RMSE. The best performance from all the runs with the various hyper-parameter configurations was an RMSE of 14,10. Using TSFEL for feature extraction resulted in a 2,28 percent improvement in LR, a 12,64 percent improvement in DT and a 0,81 percent improvement in RF RMSE. The performance of LSTM with TSFEL data resulted in very poor RMSE of 76,24. Prediction performance of the models using the TSFEL were not better than the performance of LSTM without TSFEL. The study has shown that using TSFEL for feature extraction on the NASA C-MAPSS FD001, can improve the performance of LR, RF and DT but performed very poorly on LSTM.
This study will contribute to the body of knowledge on LSTM on predictive maintenance as well as towards the use of feature extraction library TSFEL. It will also lead to future publications of academic papers in an accredited journal and conferences.