Estimation of remaining useful life of rolling element bearings based on the Adaptive Kernel Kalman filter

Abstract

Estimating the Remaining Useful Life (RUL) of Rolling Element Bearings (REBs) is a very challenging task, complicating the optimal scheduling of shutdowns and maintenance operations in industry. In recent years, a number of prognostic methodologies have been proposed, mainly categorized into 3 groups: physical model-based, AI-based, and statistical-based methodologies. Statistical based methodologies, including the Kalman filter and its variants, are commonly used in RUL estimation due to their explainability and efficiency. However, their reliance on specific Health Indicators (HIs) often restricts their generalization capabilities for different scenarios. Additionally they often face accuracy and robustness issues due to the nonlinearity of the degradation trends in rotating machines. To overcome these limitations, this study introduces a prognostic methodology based on the Adaptive Kernel Kalman Filter (AKKF), which integrates HI extraction, anomaly and failure thresholds setting, parameters estimation in the data and kernel space, and uncertainty quantification. The proposed methodology is applied to three bearing degradation datasets: an in-house dataset captured on the KU Leuven gearbox prognostics test rig, and two publicly available datasets from the University of Ferrara (UNIFE) and Xi’an Jiaotong University (XJTU). The performance of the methodology is evaluated using different metrics and is compared with the Extended Kalman Filter (EKF). The results from the aforementioned three datasets indicate that the AKKF-based methodology is promising to be used for the highly nonlinear degradation trends of REBs, achieving good results. Moreover, several issues, including the HI selection, the degradation model selection and the filter selection are discussed in detail.