Prognostic Analysis of High-Speed Cylindrical Roller Bearing Using Weibull Distribution and k-Nearest Neighbor

Abstract

Bearing remnant operational life can be determined by implementing a data-driven prognostics method. In this work, the bearing run-to-failure data from experimentation on test rig is used to extract time-domain features. The sudden change in time domain information signifies the fault inception which led to failure stage promptly. The monotonicity metric is utilized to select the optimal feature set that best represents bearing degradation. PCA (principal component analysis) is employed for dimension reduction and fusion, and a uni-dimensional health indicator (HI) is constructed. Fluctuations of HI are smoothed by fitting it with a Weibull failure rate function (WFRF) and the corresponding parameters are estimated using non-linear least squares method. By inverting the model, the predicted time values are calculated, and hence remnant operational life of bearing is evaluated and compared with the actual life from experimental data. The performance assessment metrics utilized are mean absolute percentage error (MAPE), mean squared error (MSE), root mean squared error (RMSE), and bias. Besides this, an online degradation state classification method using the k-nearest neighbor (KNN) classifier is implemented. The KNN model performance is assessed by constructing ROC (receiver operating characteristics) curve, which indicates the value of AUC (area under the curve) equal to 0.94, representing high accuracy of the KNN. The RUL (remaining useful life) is predicted within 95% confidence limits, and the predicted RUL almost follows the actual one with some fluctuations. The model performance is found promising and can be implemented to evaluate the remaining useful life of bearing.

Investigation of Machine Learning Methods for Predictive Maintenance of the Ultra-High-Pressure Reactor in a Polyethylene-Vinyl Acetate Production Process

“…Standard deviation, kurtosis, variance, peak-to-peak, and other statistical features are common for further predictions and are calculated from time-series and frequency domain data [17]. The Weibull distribution is often used to smooth oscillating statistical features over time [18]. Some also use large machine learning models such as deep convolution neural networks (CNN) to extract features from wavelet transforms [19].…”

Detecting preload degradation of a ball screw feed system using high-frequency reconstruction and support vector machine

“…State identification has been proposed to address the limitations of fault identification methods. Rathore and Harsha (2022) used the Weibull distribution and K-nearest neighbor (KNN) algorithm to predict the remaining life of a bearing. They characterized the remaining life of the bearing by looking for monotonicity (health index).…”