Identificação de desbalanceamento de máquinas rotativas usando metamodelagem cokriging

Abstract

Along with misalignment, the leading cause of mechanical issues in rotating machinery is unbalance of the rotating assembly. Among the various balancing techniques available, the influence coefficient method is widely used. However, this technique demands test runs with trial masses, which is often inefficient, resulting in the need for new balancing attempts and prolonging service execution time. In the Oil and Gas industry, minimizing intervention time is paramount, directly translating to increased productivity and reduced lost profit, the financial loss due to the non-processing of oil derivatives. Specifically for refineries (downstream), where redundant (spare) machinery is often unavailable, any unscheduled intervention can lead to a complete unit shutdown. Therefore, exploring novel possibilities for unbalance detection without the need for trial mass runs is essential to significantly reduce the time required for these procedures. Thiswork addresses this context by implementing a balancing methodology without the use of test weights using a Cokriging Metamodel. This approach leverages two distinct data sources: experimental data, which consist of time-domain vibration signals collected directly from the machine (high-fidelity or expensive data), and numerical data, which are vibration signals generated from a numerical model (low-fidelity or cheap data). To validate the proposed methodology, two validation phases were conducted: a purely numerical validation and an experimental validation. For the numerical validation, vibration signals generated from a numerical model were utilized as high-fidelity samples and Monte Carlo uncertainty was added to stiffness and damping of the bearings and rotor modulus of elasticity in order to create the low-fidelity samples. For the experimental validation, tests were performed using a test rig (a rotor kit with two disks between bearings). Unbalance sets were applied in two planes (two disks) and vibration measurements in time domain were taken. Subsequently, this vibration data was divided into training and validation datasets and a cross-validation was performed. The experimental validation is a crucial factor in establishing the methodology, for instance, defining the optimal quantity of high-fidelity and low-fidelity samples to be included in the metamodel to maximize its performance.