An attention-based stacked BiLSTM framework for predicting remaining useful life of rolling bearings

Abstract

Data-driven prognostic techniques have been widely implemented to accurately predicting the remaining useful life (RUL) of high-speed bearings. Accurate RUL estimation attributes to determining underlying relationship between bearing degradation progression and current health status. To achieve this, effective feature compression and optimal feature selection are significant challenges to be overcome. Therefore, in this paper a framework is developed which utilizes the ‘Stacked Bi-directional Long Short Term Memory’ (SBiLSTM) with attention mechanism (A-SBiLSTM) to determine the underlying relationship between the health status and RUL of bearing. Bi-directional LSTM is capable of processing the past and the future states simultaneously thereby mining the useful degradation information from time-series data. Additionally, fitness analysis is carried out using feature ranking metrics to select features highly correlated with bearing degradation. The effectiveness and feasibility of the proposed methodology are experimentally validated and generalized on dataset from PRONOSTIA platform. The comparison results based on root mean square error (RMSE), and mean absolute error (MAE) achieved superior performance compared to state-of-art methods. The experimental results revealed that the A-SBiLSTM with attention mechanism and fitness analysis achieved superior performance than other advanced methods for accurate RUL prediction.