Machine Remaining Useful Life Prediction via an Attention-Based Deep Learning Approach

TLDR

Predicting machine remaining useful life is a key prognostics task, and while deep models such as LSTM have shown strong performance, they typically rely only on the final time‑step output and may miss valuable handcrafted domain features. This study aims to combine handcrafted domain features with automatically learned features to improve RUL prediction accuracy. An attention‑based deep learning framework is proposed, where an LSTM extracts sequential features, an attention module weighs feature importance across time, and a fusion layer merges handcrafted and learned features. Experiments on two real datasets demonstrate that the proposed method surpasses existing state‑of‑the‑art approaches.

Abstract

For prognostics and health management of mechanical systems, a core task is to predict the machine remaining useful life (RUL). Currently, deep structures with automatic feature learning, such as long short-term memory (LSTM), have achieved great performances for the RUL prediction. However, the conventional LSTM network only uses the learned features at last time step for regression or classification, which is not efficient. Besides, some handcrafted features with domain knowledge may convey additional information for the prediction of RUL. It is thus highly motivated to integrate both those handcrafted features and automatically learned features for the RUL prediction. In this article, we propose an attention-based deep learning framework for machine's RUL prediction. The LSTM network is employed to learn sequential features from raw sensory data. Meanwhile, the proposed attention mechanism is able to learn the importance of features and time steps, and assign larger weights to more important ones. Moreover, a feature fusion framework is developed to combine the handcrafted features with automatically learned features to boost the performance of the RUL prediction. Extensive experiments have been conducted on two real datasets and experimental results demonstrate that our proposed approach outperforms the state-of-the-arts.