Bayesian temporal factorization and improved transformer architecture for the prediction of aero-engine remaining useful life
Chaochao Guo, Youchao SunPurpose
The purpose of this paper is to propose a comprehensive model that skillfully interpolates missing values without introducing potential bias, and performs high-precision prediction of the remaining useful life (RUL) of an engine using an improved transformer model.
Design/methodology/approach
This paper proposes a method that integrates Bayesian temporal factorization with improved transformer architecture. By integrating low-rank matrix factorization and vector autoregressive (VAR) processes into a unified probabilistic graphical model, which enables the framework to capture both global and local consistency within large-scale time series data, is essential for robust RUL prediction. Additionally, a tailored Gibbs sampling algorithm is developed to impute missing spatiotemporal engine data effectively. Following the estimation of missing values, an improved transformer model is employed for RUL prediction, which leverages gated convolutional units to enhance its ability to fuse local contextual information at each time step, and harnesses bespoke gated attention units further to improve the computational efficiency of transformer models.
Findings
Extensive experimentation on a turbofan engine dataset validates the efficacy of the proposed method. Results demonstrate that our approach either outperforms or is comparable to the best existing approaches in RUL estimation.
Originality/value
The method proposed in this study addresses the problem of poor RUL prediction capability of traditional deep learning (DL) models when engine data are partially missing.