A Deep Learning Framework Based on Novel Hierarchical-LSTM Model for Enhanced Machinery Prognostics

Abstract

Machinery prognostics has garnered increasing research attention due to its critical role in industries such as manufacturing and renewable energy. Data-driven techniques, particularly recurrent neural networks (RNNs) and convolutional neural networks (CNNs), have shown promise in accurately extracting features for estimating the remaining useful life (RUL) of machinery. However, the non-stationary and non-linear nature of machinery signals poses significant challenges to achieving accurate prognostics. This study introduces a novel hierarchical recurrent neural network method called hierarchical long short-term memory (H-LSTM) that is based on the long short-term memory (LSTM) model. H-LSTM is meant to address the problems with traditional RNNs that only use the previous time step for sequential data learning. It incorporates a hierarchical structure, enabling influence from multiple preceding time steps at each current step. Experimental evaluation on the FEMTO benchmark bearing dataset under varying operational conditions demonstrates that the proposed H-LSTM approach achieves up to fourfold improvements in performance compared to state-of-the-art methods, particularly for low signal-to-noise ratio (SNR) signals.