A Robust Health Indicator for Rotating Machinery Under Time-Varying Operating Conditions

Abstract

Bearing is an essential component whose failure leads to costly downtime in operation. Therefore, it is important to establish an accurate health indicator (HI), using which the remaining useful life can be reliably predicted. To date, most of the health assessment for bearing have been focused on the constant operating condition while in practice, it operates under various operating conditions (rotating speed and loading). Motivated by this, this paper proposes a method to extract robust HI which undergoes variable operating conditions. The idea is to cluster the operating conditions regimes, and develop HI based on the Mahalanobis distance using the optimal features subset in each regime. To validate the effectiveness, bearing run-to-fail experiment is performed under variable operating condition, and proposed HI is compared with the traditional statistical features. The remaining useful life is predicted by the data augmentation prognostics algorithm which was to overcome data deficiency problem.

Gas Turbine Anomaly Detection under Time-Varying Operation Conditions Based on Spectra Alignment and Self-Adaptive Normalization

“…Neural networks are unable to effectively match frequency components in the spectrum that occupy different positions due to varying rotational speeds, significantly impairing the learning capability of the neural network. There have been some studies dedicated to anomaly detection in time-varying operating condition equipment [25][26][27][28]. However, most of these studies have focused on bearings or gearboxes.…”

Prognosis using Bayesian Method by Incorporating Physical Constraints

“…In practice, however, it is challenging to establish highfidelity physics model for complex machinery system as the degradation process can involve various components and failure mechanisms. As an alternative, simple but effective mathematical models such as a single exponential model (Kim et al, 2022b;Wang et al, 2021), polynomial model (Kim et al, 2017) or dual-exponential model (Chen et al, 2020) are widely used in model-based approaches. However, traditional Bayesian inference-based algorithms present several challenges when used with these models, leading to inaccurate predictions of RUL.…”