Advancements in two-phase flow regime classification: a comparative study of machine learning and deep learning approaches using wire-mesh sensor data

• Ano de defesa: 2024
• Autor(a) principal: Ambrosio, Jefferson dos Santos
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Brasil; Programa de Pós-Graduação em Engenharia Elétrica e Informática Industrial; UTFPR
• Palavras-chave em Português: Máquina de vetores de suporte; Escoamento bifásico; Aprendizado profundo (Aprendizado do computador); Análise de séries temporais; Detectores; Redes neurais (Computação); Support vector machines; Two-phase flow; Deep learning (Machine learning); Time-series analysis; Sensors; Neural networks (Computer science); CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA; Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/36002

Flow regime classification is essential for analyzing and modeling two-phase flows, as it demarcates the flow behavior and influences the selection of appropriate predictive models. Machine learning-based approaches have gained relevance in flow regime classification research in the last few years. However, they are still solidly based on the construction and careful definition of hand-crafted features. Deep learning approaches, on the other hand, can provide more robust and end-to-end solutions. However, they are underexplored and have not evaluated the generalization of the models to other data or acquisition systems. In this thesis, we compare two different approaches for classifying flow patterns (churn, bubbly, and slug) using time series of void fraction from a wire-mesh sensor. In the first, defined as MoG+SVM, the time series is modeled as a stochastic process of independent and identically distributed samples with probability density function described by a Mixture of Gaussian (MoG) model. The estimated parameters of the mixture are then fed into a Support Vector Machine (SVM), yielding the flow pattern classification. The second, defined as SOTA-DL, we propose using end-to-end state-of-the-art (SOTA) time-series classification methods (ResNet, LSTM-FCN, and TSTPlus) for two-phase flow patterns. We also present the generalization analysis of the models with cross-dataset experiments, training the model with one dataset and testing it with another dataset collected in another system for two datasets, here defined as HZDR and TUD. The results demonstrate that the deep learning-based approach (SOTA-DL) presents superior classification metrics in all cases evaluated, particularly in cross-dataset experiments. With the proposed SOTA methods, all the evaluated metrics (accuracy and F1-Score) consistently surpass 90% in all cases, while the MoG+SVM method can decrease the performance under 80%. This demonstrates the relevance of the analysis proposed here for flow regime classification literature and opens up a new set of possibilities for research in this area, aiming at robust solutions that are viable for practical use.