Predicting remaining useful life using spectral normalized LSTM encoder and Gaussian processes

Remaining Useful Life (RUL) is the measure of the expected functional duration of the certain component or system (Liu et al., 2020). This measure is used in several application fields, such as industrial equipment, as in the case of this work. An accurate prediction of RUL is of high interest, since it can improve the decision-making in these areas, optimize costs and ensure safety in critical applications. To improve the reliability of the predictions, some works highlight the necessity of incorporating uncertainty to the estimation of RUL. In this project a deep model based on a Long Short-Term Memory combined with a Variational Gaussian Process regression was used for the prediction of RUL of CMAPSS dataset (Saxena et al., 2008) aerospacial engines, as well as its related uncertainty. The performance of the model was assessed against benchmark literature works that followed similar processes. Although the performance of the model for both point prediction and uncertainty assessment is worse than common benchmark metrics, it still offers competent results in the low RUL region. We state that this performance might be improved by applying self-Attention mechanisms to the encoder architecture, as well as considering uncertainty prediction in both stages of training (LSTM and GP).