Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
FAULT ANALYSIS IN A ROTOR SUPPORTED BY ROLLER BEARINGS USING CLUSTERING TECHNIQUES
Submission Author:
Gustavo Storti , SP
Co-Authors:
Nathali Dreher, Iago Almeida, Gustavo Storti, Gregory Bregion Daniel, Tiago Machado
Presenter: Nathali Dreher
doi://10.26678/ABCM.COBEM2021.COB2021-0594
Abstract
The interest in fault detection in rotating systems has been studied for a long time, but the increasing development related to machine learning in recent years has intensified researches on application of these algorithms to establish efficient and automatic forms for this purpose. In rotor dynamics, the bearings are a key element to understand the system behavior. Rotating machinery maintenance is performed to avoid unscheduled stops, which usually leads to productive and financial losses. To mitigate these losses, health monitoring of these systems is vital. However, the techniques used to do so are not always easy to apply in an industrial environment and may require extensive experience and study. In this paper, fault detection on bearings through unsupervised machine learning algorithms is investigated. Three different clustering methods are applied using the experimental vibration signals of a rotor in which the electric motor is supported by roller bearings operating with different fault locations (ball, inner race, and outer race) and damage severity. The applied clustering methods are k-means, DBSCAN, and hierarchical clustering, very known and widely used methods. The vibration signals were provided by the Case Western Reserve University Bearing Data Center, which is a standard reference to test fault detection algorithms. The clustering is based on time-domain parameters extracted from the vibration signals. In this work, the results are divided into three categories to analyze the effectiveness of the methods: by severity, location, and a case in which all data are tested. It is expected to compare the efficiency of clustering methods by an appropriate metric of accuracy to recognize the database based on the number of defects and severity just by evaluating the rotor vibration signal. The number of clusters, an input in some of these algorithms, is determined by three metrics, namely Silhouette, Davies-Bouldin, and Elbow. All the metrics tend to produce similar results, providing the approximate number of clusters of the data set. The expected potential for the data segmentation in different conditions and the small interaction required to employ unsupervised learning techniques indicate this approach as a favorable tool for real applications.
Keywords
rotating machinery, clustering algorithms, unsupervised learning, bearing fault analysis
