Abstract:

This study investigates the use of Long Short-Term Memory (LSTM) for Remaining Usable Life (RUL) prediction from the Jet Engine Simulated Dataset (C-MAPSS) engine dataset and the impact and contribution of different data pre-processing techniques on this prediction. After various data normalization techniques, the dataset is filtered using Savitzky-Golay filtering, wavelet transform and exponential moving average (EMA). Each filtering technique, together with the normalization methods, is applied to the data set separately and the effectiveness of the LSTM model in predicting RUL is evaluated for each combination. Quantitative analysis of the experimental results shows that appropriate normalization and filtering strategies applied to time series data improve the training phase of the LSTM models, thereby increasing the accuracy of RUL prediction. In this study, it is shown that the choice of the best data pre-processing structure will directly affect the efficiency of network training and thus it is possible to optimize RUL prediction with the LSTM model.

Keywords: