Remaining Useful Life Prediction of Rolling Element Bearings Based on Unscented Kalman Filter

Abstract

A data-driven methodology is considered in this paper focusing towards the Remaining Useful Life (RUL) prediction. Firstly, diagnostic features are extracted from training data and an analytical function that best approximates the evolution of the fault is determined and used to learn the parameters of an Unscented Kalman Filter (UKF). UKF is based on the recursive estimation of the Classic Kalman Filter (CKF) and the Unscented Transform, presenting advantages over the Extended Kalman Filter (EKF) for high non-linear systems. The learned UKF is further applied on testing data in order to predict the RUL under different operating conditions. The influence of the starting point of the prediction is analyzed and a method for the automated parameter tuning of the Kalman Filter is considered. In the end, the result is evaluated and compared to CKF and EKF on experimental data based on dedicated performance metrics.

A Framework for Bearing Prognostics in Changing Operating Conditions

“…In their paper Qi et al [1] propose a novel spectral based entropy condition indicator and use a particle filter to estimate the parameters of an exponential model used to fit the indicator. The RUL is estimated at each time step by systematically progressing the condition indicator to a threshold marking end of life using the exponential model.…”

“…Bearings fail due to a wide number of reasons including lack of/insufficient lubrication, lubricant contamination, corrosion and misalignment. It is therefore important to not only detect and identify bearing faults but also be able to accurately predict remaining useful life (RUL), thus determining the best maintenance opportunities [1]. Two of the most common approaches in predicting RUL are either to use a physics based model or a data driven one.…”

“…the root mean square (RMS), kurtosis and skewness are commonly employed in bearing diagnosis and prognosis [4]. Unfortunately, these condition indicators may fail to exhibit crucial characteristics such as monotonicity in non-simulated data because it is characterized by high noise, low impulsiveness and changing operating conditions [1].…”

“…Huang et al [5] developed a Kurtosis based condition indicator and used a convolutional long short-term memory to predict both the indicator and RUL with very accurate results. Qi et al [1] developed an entropy based condition indicator and used a particle filter to fit the parameters of an exponential degradation model. The authors were able to forecast the indicator and obtain RULs within a ±20% error band.…”

A scale independent flexible bearing health monitoring index based on time frequency manifold energy & entropy

“…Constructing an appropriate health monitoring index to identify bearing health condition with precision is an important aspect of prognosis for rolling element bearings. Moreover, in order to perform proper health stage division and the prediction of remaining useful life, the aim of any health monitoring index is to be as sensitive as possible to the determination of the timeto-start prediction (TSP) point, which is the point where the first signs of abnormality in the bearing are detected, indicating the potential performance degradation of the bearing [4][5][6]. From this perspective, entropy is a promising tool for the health characterization of rolling element bearings [7,8].…”