Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
Machine Learning Techniques to Fault Detection in Rotating Machines
Submission Author:
Luís Otávio Garavaso , SP , Brazil
Co-Authors:
Luís Otávio Garavaso, Gregory Bregion Daniel, Katia Lucchesi Cavalca Dedini
Presenter: Luís Otávio Garavaso
doi://10.26678/ABCM.COBEM2021.COB2021-0634
Abstract
Rotating machines are necessary for power generation since they perform vital roles that extend from extracting resources, such as fuels, to converting kinetic energy from water and airflow into electricity for consumption. These machines, however, face different types of mechanical faults that alter the system's vibration response creating specific patterns known as fault signatures. Therefore, interpreting these vibration marks with adequate tools leads to proper fault identification, which improves maintenance scheduling, reduces repair time and machinery breakage. While some of these signatures can be associated to the first harmonic response, such as the rotating unbalance affecting the amplitude of the measured signal, others are still not fully understood, which turns Machine Learning into a powerful tool for fault diagnostics. This research, therefore, combines Mechanical Engineering knowledge with Machine Learning techniques to train widely used algorithms, such as Logistic Regression, Support Vector Machines, and Artificial Neural Networks, to identify through vibration data the existence of mechanical faults in rotating machines. The data is initially generated through numerical integration of equations of motion added by different levels of noise. Rotor modeling, otherwise, can be analytical, as in the case of Laval Rotor lumped parameters solution, or through the Finite Elements Method, for more complex geometries. The Support Vector Machine and the Artificial Neural Networks have shown to be reliable algorithms for rotating unbalance detection. However, it is necessary to acknowledge the Support Vector Machine's performance for requiring way less data than the Artificial Neural Networks. Finally, this work is promising to diagnostics techniques evaluation depending on the fault signature of rotating machines.
Keywords
Rotordynamics, machine learning, Support vector machine, artificial neural network, Fault Identification
