Balanceamento de máquinas rotativas utilizando redes neurais

• Autor(a) principal: Cabral, Leonardo Dias da Silva
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFU
• Texto Completo: https://repositorio.ufu.br/handle/123456789/38769; http://doi.org/10.14393/ufu.di.2023.377
• topic: Dinâmica de rotação; Balanceamento de máquinas rotativas; Rede neural; Ausência de massa de teste; Rotation Dynamics; Balancing of Rotating Machines; Neural networks; Absence of Test Mass; Engenharia mecânica; CNPQ::ENGENHARIAS::ENGENHARIA MECANICA::MECANICA DOS SOLIDOS; Engenharia mecânica; Redes neurais (Computação); Balanceamento de máquinas; Dinâmica das máquinas

As studies and needs have progressed over the years, methodologies have been proposed to minimize one of the main causes of vibration observed in rotating machines, which is unbalanced. Balancing aims, in general, to equalize the mass distribution along the rotor, which allows reducing the forces induced by such irregularities and, consequently, their vibration responses. Thus, balancing provides a better operating condition for the machine, with improved performance and compliance with safety limits specified by technical standards. The balancing techniques developed over time have diversified into alternatives that seek to utilize the relationship between the observed dynamic behavior of the rotating machine and its physical and geometric characteristics, in order to predict the correction masses and their respective positions on the rotor. Among these techniques, it is possible to highlight more traditional approaches such as the Influence Coefficients Method (which utilizes trial weights) and more recent ones, with the advent and popularization of artificial intelligence techniques. This work proposes an approach to balance rotating machines without using trial weights, employing neural networks. In general, the proposed neural network aims to establish a relationship between the vibration responses obtained in the machine and the possible correction masses and positions through training with a previous database. This work addresses six configurations, where the first four use the finite element method () to form the database (as a numerical test). The last two configurations are experimental in nature (i.e., the network's database comes from experimental bench data). This dissertation also proposes a data expansion approach that reduces the amount of unbalance conditions required for network training, which is used for numerical and experimental neural network configurations. Each neural network configuration is implemented and evaluated, and the results show promise.