Three-stage feature selection approach for deep learning-based RUL prediction methods

Abstract

The remaining useful life (RUL) prediction plays an increasingly important role in predictive maintenance. With the development of big data and the Internet-of-Things (IoT), deep learning (DL) techniques have been widely adopted for RUL prediction. Addressing the limitation of the current methods for data under multiple operating conditions, this paper proposes a three-stage feature selection approach for DL-based RUL prediction models. The k-medoids cluster is initially used to sort raw data based on different operating conditions. In the first stage of feature selection, an operational-based normalisation approach is applied to reconstruct the data. Afterwards, Spearman's rank and pair-wise Pearson correlation coefficients are used to eliminate irrelevant and redundant features in the second and third stages, respectively. A case study using NASA's Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset is presented to quantitatively evaluate the influence of the proposed feature selection method using the Recurrent Neural Network (RNN) and its’ variants, enhanced by an optimised activation function and optimiser. The results confirm that the proposed method can improve the stability of DL models and achieve about a 7.3% average improvement in the RUL prediction for popular and state-of-the-art DL models.