A hybrid physics-informed neural network for main bearing fatigue prognosis under grease quality variation. (15th May 2022)

• Record Type: Journal Article
• Authors: Yucesan, Yigit A.; Viana, Felipe A.C.
• Is Part Of: Mechanical systems and signal processing. Volume 171(2022)
• Publication Date: 2022-05-15
• Subjects: hybrid physics-informed neural network -- Applied machine learning -- Wind turbine bearing fatigue -- Uncertainty quantification; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2022.108875
• Languages: English
• ISSNs: 0888-3270

Abstract

Abstract: Fatigue life of a wind turbine main bearing is drastically affected by the state of the grease used as lubricant. Unfortunately monitoring the grease condition through predictive models can be a daunting task due to uncertainties associated with degradation mechanism and variations in grease batch quality. Eventually, discrepancies in the grease life predictions caused by variable grease quality may lead up to inaccurate bearing fatigue life predictions. The convoluted nature of the problem requires a novel solution approach; and in this contribution, we propose a new hybrid physics-informed neural network model. We construct a hybrid model for bearing fatigue damage accumulation embedded as a recurrent neural network cell, where reduced-order physics models used for bearing fatigue damage accumulation, and neural networks represent grease degradation mechanism that quantifies grease damage that ultimately accelerates bearing fatigue. We outline a two-step probabilistic approach to quantify the grease quality variation. In the first step, we make use of the hybrid model to learn the grease degradation when the quality is the median of the distribution. In the second step, we take the median predictor from the first step and track the quantiles of the quality distribution by examining grease samples of each wind turbine. We finally showcase our approach with a numerical experiment, where we test the effect of the random realizations of quality variation and the number of sampled turbines on the performance of the model. Results of the numerical experiment indicate that given enough samples from different wind turbines, our method can successfully learn the median grease degradation and uncertainty about it. With this predictive model, we are able to optimize the regreasing intervals on a turbine-by-turbine basis. The source codes and links to the data can be found in the following GitHub repository https://github.com/PML-UCF/pinn_wind_bearing . Highlights: Hybrid physics-informed neural network framework for cumulative damage modeling. Wind turbine main bearing fatigue failure prognosis. Probabilistic approach for grease quality uncertainty quantification.