Bayesian Dual-Input-Channel LSTM-Based Prognostics: Toward Uncertainty Quantification Under Varying Future Operations

Deep learning methods have received tremendous attention in remaining useful life (RUL) prediction in recent years. Despite the promising results achieved by those deep learning methods, they failed to take account of the impact of varying future operations on RUL prediction and prognostics uncertainty. In most industrial scenarios, the RUL of products is closely related to future missions and loading profiles, and they ought to be considered for RUL prediction. In addition, Bayesian deep learning treats the model parameters as random variables, and takes advantage of the Bayesian formulas for adaptive model parameter updating to obtain credible intervals of RUL prediction. In this article, a Bayesian dual-input-channel long short-term memory (BDIC-LSTM) network is put forth to conduct the point estimation and credible interval estimation of RUL prediction. The BDIC-LSTM network consists of a DIC-LSTM network and an improved Monte Carlo dropout (IMCD) method to effectively extract the features of future operations and quantify the prognostics uncertainty, respectively. In the DIC-LSTM network, the raw signal data are fed into the main input channel consisting of LSTM modules. Meanwhile, bidirectional LSTM (Bi-LSTM) modules are leveraged as an auxiliary input channel to fully extract the future operation information in the operation data. The IMCD method is investigated to estimate the credible intervals of RUL via decomposing the prognostics uncertainty into aleatory uncertainty of measurement data and epistemic uncertainty of the network. For the training of the BDIC-LSTM network, a padding and packing training mode, an improved loss function, together with a Lookahead optimizer, are devised to accelerate the convergence speed and enhance the accuracy of prognostics. Experiments on the C-MAPSS dataset are carried out to validate the effectiveness of the proposed method.

I. Introduction

Prognostics and health management (PHM) has gained widespread attention in both industry and academia, owing to its ability to reduce system downtime while simultaneously boosting system safety and reliability. Remaining useful life (RUL) prediction, an essential technology in PHM, can offer the expected residual time and health status of a system based on massive raw signal data [1]. An accurate and trustworthy RUL prediction result is not only an essential element of preventing systems’ unexpected incidents, but also a basis for future predictive maintenance (PdM) of the system. As a result, a plethora of effective models and tools have been developed to extract degradation features and predict RUL for many industrial applications to ensure their safety against failures and save operation and maintenance costs [2], [3].