Modelling of Journal Bearings for Predictive Maintenance

Abstract

Large medium voltage machines such as motors and generators, running with a low speed of less than 1000 rpm, are typically designed with the use of sleeve (or journal) bearings. These machines play a critical role in industrial processes. Sleeve bearings are simple in construction yet designed to operate for many years without any maintenance. Since these are critical components of rotating machines, knowledge about their condition is fundamental. Typical and well-proven methods for condition monitoring of journal bearings are based on measurement of the shaft movement within lubrication oil or monitoring the condition of lubrication oil itself. Both techniques require the installation of special additional sensors that are typically very costly and not necessarily feasible for the systems already in operation. Instead, this article proposes to use existing large data sets of performance-related measurements from rotating machines equipped with sleeve bearings and model them in order to detect anomalies, preferably originating from potential bearing faults. The aim of modelling is to predict bearing temperature as it impacts physically and predictably lubrication oil viscosity and thus lubrication quality. Models derived from both linear and non-linear approximations are to be benchmarked. Since at this stage of analysis, the training process is unsupervised (due to lack of labels for confirmed bearing fault), recommendations given in the article are fundaments for a follow-up work aiming at enriching the data with simulated or field-confirmed bearing defects or suspensions.