Balanceamento de máquinas rotativas sem o uso de massas de teste através de redes neurais artificiais

Abstract

Among the most common faults in rotating machines, imbalance stands out as the most frequently encountered. It is characterized by an uneven distribution of mass along the axis of rotation and is inherent to this type of equipment, whether due to imperfections in manufacturing and assembly processes or wear caused by prolonged use. Thus, to ensure operational safety and extend equipment lifespan, the periodic application of balancing procedures becomes necessary. Over time, several balancing methodologies have been developed, with the Influence Coefficient (IC) method being widely adopted in industry. However, with increasing demand for more efficient machines, the machine downtime required to perform balancing using the IC method has become increasingly inconvenient. In this context, this Thesis presents two methodologies for rotor balancing without the use of trial masses. The first, model-based methodology employs a finite element model of the rotor to generate training data for an Artificial Neural Network (ANN). This approach is validated numerically and experimentally, and the adequacy of the rotor model representation is analyzed in comparison with the estimates provided by the neural model. The second, data-based methodology introduces a novel Data Augmentation (DA) approach, specifically developed to generate virtual samples from a limited set of experimental data. The results show that the first methodology proved robust under significant variations in parameters considered uncertain in the model, achieving an average reduction of 50% in vibration amplitudes across three case studies. Regarding the second methodology, two case studies were conducted. In the first, five experimental samples were used, resulting in an average reduction of 60% in vibration amplitudes. In the second study, only two experimental samples were employed, yielding an average reduction of approximately 70% across all tested conditions. Therefore, both methodologies prove to be viable solutions for balancing rotating machinery without the use of trial masses — the first being recommended when a numerical model of the rotor is available, and the second when no such model exists and there is a limited amount of experimental data.