A rotordynamic analysis of hole-pattern gas annular seals using a two-control-volume model, Ha and Childs and frequency dependent transfer-function model, Kleynhans and Childs is modified with four features. The energy equation is added, and real gas properties are used instead of the ideal gas equation of state. The depth of the hole-pattern is made variable with the axial distance along the seal. And last, the addition of deep grooves to hole-pattern seals is analyzed, and the code’s predictions for the influence of a groove are compared with test data.

Results are presented for experimental and theoretical rotordynamic coefficients and leakage of straight smooth annular gas seals. Experimental rotordynamic coefficients were measured and trends in changes of rotordynamic coefficients with operating variables such as rotor speed, back-pressure, fluid preswirl, and seal clearance are analyzed. Experimental results show that cross-coupled stiffness coefficients are highly influenced by fluid preswirl and only moderately influenced by other operating parameters, whereas direct damping is nearly unaffected by changes in operating parameters. Effective damping, a good indicator of stability, is highly affected by fluid preswirl. Although rotordynamic coefficients of straight smooth annular gas seals are assumed to be frequency independent, experimental results suggest a frequency dependent nature at high back-pressures and high excitation frequencies. Experimental results for rotordynamic coefficients and leakage are compared with theoretical predictions of ISOTSEAL, an isothermal-flow, two-control-volume, bulk-flow rotordynamic analysis program. All rotordynamic coefficients are underpredicted. Direct stiffness is poorly predicted while cross-coupled stiffness and direct damping are predicted reasonably well. Leakage is also consistently under-predicted. Theory predicts a slight frequency dependent nature for a limited number of test configurations.

A rotordynamic analysis of hole-pattern gas annular seals using a two-control-volume model, Ha and Childs and frequency dependent transfer-function model, Kleynhans and Childs is modified with four features. The energy equation is added, and real gas properties are used instead of the ideal gas equation of state. The depth of the hole-pattern is made variable with the axial distance along the seal. And last, the addition of deep grooves to hole-pattern seals is analyzed, and the code’s predictions for the influence of a groove are compared with test data.

[pt] A correta operação de equipamentos na indústria de Óleo e Gás é fundamental para a reduzir perdas ambientais, humanas e financeiras. Neste cenário, foram estudados selos secos a gás (em inglês,DGS) de compressores cetrífugos, por serem identificados como os mais críticos devido à extensão dos danos potenciais causados em caso de falha. Neste estudo, foram desenvolvidos 31 modelos regressivos disponíveis no Scikit-Learn através de técnicas de aprendizado de máquina (em inglês, ML). Estes foram treinados com um conjunto de dados escassos, criado a partir de uma técnica de planejamento de experimentos, para substituir simulações numéricas na previsão de confiabilidade operacional de DGSs. Primeiramente, foi validado um modelo baseado na simulação da Dinâmica dos Fluidos Computacionais (em inglês, CFD) para representar o escoamento do gás entre as faces de selagem, a fim de possibilitar o cálculo da confiabilidade operacional do equipamento. Neste, foi utilizado o software de CFD de código aberto OpenFOAM em conjunto com o banco de dados de substâncias do software REFPROP, a fim de possibilitar ao usuário definir a mistura gasosa e as condições operacionais avaliadas. Em seguida, foram realizados dois estudos de caso seguindo um fluxograma genérico de projeto proposto. O primeiro consistiu na determinação de um modelo regressivo para estimar a confiabilidade de um DGS cuja composição gasosa (composta por metano, etano e octano) é fixa porém suas condições operacionais podem ser alteradas. Já o segundo consistiu na determinação de um modelo regressivo mais robusto, onde tanto a composição gasosa como as condições operacionais podem ser alteradas. Por fim, foi avaliada a viabilidade de implementação de ambos os modelos em condições reais de operação, baseado na norma infinita obtida para a predição do conjunto de teste. As performances atingidar foram de 1.872 graus Celsius e 6.951 grau Celsius para o primeiro e segundo estudos de caso, respectivamente. / [en] The correct equipment operation in the Oil and Gas industry is essential to reduce environmental, human, and financial losses. In this scenario, dry gas seals (DGS) of centrifugal compressors were studied, as they are identified as the most critical device due to the extent of the potential damage caused by their failure. In this study, 31 regression models available at Scikit-Learn were developed using machine learning (ML) techniques. They were trained with a scarce dataset, created based on a design of experiment technique, to replace numerical simulations in predicting the operational reliability of DGSs. First, a model based on Computational Fluid Dynamics (CFD) simulation was validated to represent the gas flowing between the sealing faces, to enable the calculation of the equipment’s operational reliability. Thus, the open-source CFD software OpenFOAM was used together with the substance database of the software REFPROP, to allow the user to define the gas mixture and the evaluated operational conditions. Then, two case studies were carried out following a proposed generic workflow. The first comprised determining a regression model to estimate the reliability of a DGS whose mixture composition (composed of methane, ethane, and octane) is fixed but its operating conditions can vary. The second consisted of determining a more robust regressive model, where both the mixture composition and the operational conditions can vary. Finally, the feasibility of implementing both models under realistic operating conditions was evaluated, based on the infinity norm obtained for the prediction of the test set. The performances achieved were 1.872 degrees Celsius and 6.951 degrees Celsius for the first and second case studies, respectively.