克里金
核(代数)
高斯过程
计算机科学
回归
过程(计算)
核回归
人工智能
模式识别(心理学)
数据挖掘
不确定度量化
机器学习
高斯分布
算法
数学
统计
操作系统
组合数学
物理
量子力学
作者
Jin Cui,Jinchen Ji,Tianxiao Zhang,Qing Ni,Licai Cao,Zixu Chen
标识
DOI:10.1177/14759217241282876
摘要
Accurate prediction and uncertainty quantification (UQ) of bearings’ remaining useful life (RUL) are essential for the safe operation of critical machinery. Conventional machine-learning-based UQ methods for RUL prediction involve specialized designs and multiple inference iterations, reducing prognostic efficiency in engineering applications. Accordingly, this study presents a novel sparse Gaussian process regression (GPR) method that uses a time-aware spatiotemporal kernel to predict bearing RUL with UQ. The architecture of the proposed approach primarily consists of two components. The first is a dual-branch network used to extract deep features from input sequences. One branch utilizes a residual convolutional gated recurrent unit to extract deep spatiotemporal features from input sequences, while the other branch is an encoder that processes time information to enrich feature variability over time. The second component is the RUL prediction segment with sparse GPR, which offers not only accurate RUL estimations but also effective UQ from the acquired deep features. The proposed method facilitates systematic and concurrent training optimization of these two components, achieving efficient and accurate end-to-end RUL prediction. The proposed method was validated on two experimental datasets. Compared to existing approaches, it enhances RUL prediction accuracy and provides effective UQ, facilitating the development of reliable maintenance strategies.
科研通智能强力驱动
Strongly Powered by AbleSci AI