A generalised machine learning model based on multinomial logistic regression and frequency features for rolling bearing fault classification

Abstract

Intelligent fault classification of rolling element bearings (REBs) using machine learning (ML) techniques increases the reliability of industrial assets. One of the main issues associated with ML model development is the lack of training data and, most importantly, the ability of models to be used for applications without specific training data, ie the generalisation capability of models. This study investigates the feasibility of using multinomial logistic regression (MLR) as generalised ML models for rolling element bearing fault classification without the requirement of training data for new bearing designs and varied machine operations. This has been achieved by using bearing characteristic frequencies (BCFs) as inputs to the MLR models extracted by a newly developed hybrid method. The new method combines cepstrum pre-whitening (CPW) and full-band enveloping, which can effectively identify the BCFs in vibration data from various machines. This paper presents the methods of the feature extraction and the development of generalised ML models for REBs based on data from the EU Clean Sky 2 I2BS project1. This model is then validated by data from Case Western Reserve University (CWRU) and the Society for Machinery Failure Prevention Technology (MFPT), available in the public domain without further training.