Predictive Bearing Maintenance Based on Transfer Learning with Preprocessing and Machine Learning Models Analysis

Abstract

In energy and power systems, the bearing is a crucial part of machineries such as generators and motors. The analysis of preprocessing methods and machine learning models is presented through validating bearing conditions classification. Two types of bearing (drive end and fan end) conditions are obtained from Case Western Reserve University Bearing Data Center, in which the drive end condition is used to train a basic model. The corresponding model is then used to evaluate the fan end condition (namely transfer learning). The features for machine learning are generated from a series of preprocessing: pre-normalization, envelope, skewness, kurtosis, root mean square, standard deviation, Fourier transform, DC removal, post-normalization, and frequency-domain features extractions. Repeatable preprocessing and machine learning algorithms are explored. Numerical preprocessing methods for time-domain and frequency-domain feature extractions are suggested. The model could predict faults from different locations using data from only a single location.