Eventos Anais de eventos
DINAME 2023
XIX International Symposium on Dynamic Problems of Mechanics
Adapting Deep Neural Networks for Rotating Machine Balancing Without Employment of Trial Weights
Submission Author:
Leonardo Cabral , MG
Co-Authors:
Raimundo Neto, Leonardo Cabral, Arinan Dourado, Aldemir Ap Cavalini Jr, Valder Steffen Jr
Presenter: Leonardo Cabral
doi://10.26678/ABCM.DINAME2023.DIN2023-0075
Abstract
During normal operation of rotating machines, irregular mass distribution along the rotor yields unbalancing forces, causing high vibration responses significantly affecting the system's health and safe operation. Hence, balancing procedures are periodically applied to rotating machines to reduce vibration amplitudes, thus keeping the system operating within acceptable safety limits. Over the years, various methods have been developed to balance rotating machines, usually relying on the use of trial weights; the assumption of linearity between unbalance forces and measured vibration; or the use of high-fidelity models of the rotating system. In this contribution, we proposed a novel balancing procedure that does not employ trial weights or require the use of a representative model of the rotating system. Here, based on scarce data sets derived from distinct operational conditions, deep neural networks are trained to model the unknown relationship between unbalanced forces and correction masses. To illustrate the capabilities of the proposed methodology, a numerical case study is presented. In the presented numerical example, an unbalanced rotor finite element model combined with an established balancing procedure (the coefficient of influence method) is used to generate a reduced data set for the considered neural networks. A such example illustrates the use of available historical balancing data from the rotor system to train the proposed artificial neural networks. Additionally, a convergence analysis is also presented, evaluating the number of unbalanced responses required to train the artificial neural networks to achieve satisfactory vibration reduction. Obtained results show that with relatively small data sets, the proposed methodology can achieve low levels of vibration, without requiring the use of trial weights, representative models, or the assumption of linearity.
Keywords
rotor balancing, neural networks, no trial weights, adapting deep neural networks, Rotating machine
