Predictive maintenance can prevent machine failure by estimating the remaining useful life (RUL) in advance. In the literature, various data-driven methods that require labelled data cannot be useful when the failure data is scarce. In this paper, an unsupervised deep-learning framework is proposed that consists of (i) an autoencoder for reducing the dimensionality of features into a virtual health index (VHI), (ii) a Bi-LSTM-based encoder-decoder with an attention mechanism and a sliding window method for predicting RUL using VHI dataset, and (iii) K-means clustering and a neural network for estimating the health status using VHI dataset. The proposed framework is tested on one of the C-MAPSS datasets, and the results are compared with the literature. Results demonstrate satisfactory performance by the proposed framework in estimating accurate RUL and health states for many turbofan engines.
Sharma et al. (Wed,) studied this question.