A Novel Remaining Useful Life Prediction Approach Combined eXtreme Gradient Boosting and Multi-Quantile Recurrent Neural Network

Abstract

Remaining useful life (RUL) prediction is a key technology to ensure the reliability and safety of high-end equipment. Deep learning is widely used for RUL prediction due to the excellent feature extraction ability and nonlinear fitting ability. Traditional recurrent neural networks adopt recursive strategy, which easily lead to the problems of error accumulation and low stability. On the other hand, most deep learning methods are used for point prediction and cannot quantify uncertainty in the prediction results. Although some probability prediction methods based on deep learning can provide probability prediction results, it require the assumption that the prediction results follow a specific distribution in advance. However, the distribution of most prediction results does not match a suitable distribution function. To solve above problems, a novel RUL prediction approach based on eXtreme gradient boosting (XGBoost) and multi-quantile recursive neural network (MQ-RNN) is proposed in this article. Specifically, XGBoost is used to select features closely related to RUL, and the selected features are fed into MQ-RNN to train the RUL prediction model. The advantage of MQ-RNN is that it has non-parametric framework, prediction results can be obtained by multi-quantile regression, which does not require prior assumptions about the distribution of predicted results. Furthermore, the proposed framework is verified by C-MAPSS dataset. Finally, comparative experiments are conducted. The experimental results show that the proposed method maintains good predictive performance in both point estimation and interval estimation.

Keywords

• Remaining useful life
• eXtreme gradient boosting
• multi-quantile recursive neural network
• quantile regression
• forking sequences training scheme