The LPST-Net: A new deep interval health monitoring and prediction framework for bearing-rotor systems under complex operating conditions

Abstract

Accurate and science-based prediction of bearing performance degradation has been a principal concern and a critical challenge issue in the sector of Prognostics and Health Management in the industry as the solution to promote the engineering system's reliability, availability, and maintainability. Deep Learning (DL) methods are currently a hotspot in Prognostics and Health Management. Nonetheless, with complex operating conditions, existing forecast models still inevitably suffer from three fatal flaws. Firstly, dynamic modern industrial systems and harsh operational environments make the degradation data of the bearing-rotor system highly stochastic and nonlinear. Secondly, in practice, the bearing-rotor system failure process will be governed by complex failure dynamics and degradation mechanisms, resulting in degradation data characterized by intense temporal order. Thirdly, deep learning models use a multi-layer or cellular design containing numerous weights and biases, generating more computational overhead. To fill this research gap, a new deep interval health monitoring and prediction framework named Lightweight Probabilistic Spatiotemporal (LPST-Net) is proposed, which is integrated with the concepts of lightweight and interval prediction and is capable of state monitoring and Remaining Useful Life (RUL) prediction for bearing-rotor systems under complex operating conditions. Mainly inspired by the improvement of the Gate Recurrent Unit (GRU), this paper designs a time series variable prediction algorithm and derives a new formulation named Weight Diminish Recurrent Unit (WDRU). It dramatically reduces the training parameters of the proposed LPST-Net framework and improves the convergence speed while ensuring prediction accuracy. The degradation data are obtained under 2 actual and complex operating conditions of bearing-rotor system unbalance and high temperature. The three metrics show that the proposed LPST-Net framework can achieve high-precision point prediction, suitable prediction intervals, and reliable probabilistic prediction results. It is also verified that the proposed LPST-Net framework has superior performance and more practical application value compared with seven mainstream methods, such as (b)Squeeze-WDRU-GPR-Net.