@Article{sdhm.2026.084325,
AUTHOR = {Xinyu Jia, Qingjie Wang, Rongchuan Zhang, Heng Ouyang},
TITLE = {A Bayesian Tree-Structured Kernel Learning Approach for Gaussian Process Prognostics of Ship Propulsion Systems},
JOURNAL = {Structural Durability \& Health Monitoring},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/sdhm/online/detail/27825},
ISSN = {1930-2991},
ABSTRACT = {Accurate state assessment and remaining useful life (RUL) prediction for ship propulsion systems are critical in ensuring maritime safety, reducing maintenance costs, and enhancing transportation network resilience. Although Gaussian process methods are widely used for RUL prediction, existing models often rely on deterministic assumptions or fixed kernel functions, which struggle to fully capture the nonlinear, multi-scale, and stochastic characteristics of degradation in ship propulsion components. This paper proposes an automatic kernel selection framework based on Bayesian tree structures, integrated with Gaussian Process Regression (GPR), for probabilistic degradation modeling and RUL prediction of critical components. The method uses additive and multiplicative operators to combine interpretable base kernels, thereby constructing a composite kernel space. A Bayesian inference strategy is employed to search for kernel structures and estimate hyperparameters simultaneously. To efficiently explore the kernel space, a two-stage beam search algorithm using model evidence is designed, along with a validation-assisted optimization mechanism to improve prediction accuracy and uncertainty calibration. Experiments were conducted on the PRONOSTIA bearing dataset and a high-fidelity ship propulsion system simulation dataset. The results show that the proposed method achieves competitive prediction accuracy and provides more reliable uncertainty quantification than fixed-kernel GPR models, supporting its application in health monitoring and risk-aware predictive maintenance of ship propulsion systems.},
DOI = {10.32604/sdhm.2026.084325}
}