Evaluating machine learning techniques for detection of flow instability events in offshore oil wells

• Ano de defesa: 2021
• Autor(a) principal: Carvalho, Bruno Guilherme
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Mestrado em Informática; Centro Tecnológico; UFES; Programa de Pós-Graduação em Informática
• Palavras-chave em Português: Instabilidade de fluxo; aprendizado de máquina; validação cruzada; seleção de características; algoritmo genético; subject.br-rjbn; Ciência da Computação
• Link de acesso: http://repositorio.ufes.br/handle/10/15481

Flow instability is an abnormal operational state in offshore oil wells. For the oil and gas industry, methods to detect and classify faults as soon as possible are crucial to reduce downtime and increase efficiency. The application of machine learning algorithms has been extensively applied in an industrial context, proven to be a viable way to tackle this kind of problem. In this study, an evaluation is performed on the application of machine learning techniques for the detection and classification of pressure and temperature sensor readings related to flow instability. Firstly, a custom cross-validation splitting strategy is defined and compared to the classical equal split. Results are shown to be much more realistic when checked on previous publications. Next, grid search is chosen to evaluate whether hyperparameter tuning could increase the classifier’s performance. Results were not satisfactory. Then, feature selection is applied to reduce problem dimension and circumvent the curse of dimensionality. Three different methods were used: sequential feature selection, hybrid ranking wrapper, and genetic algorithm. Only a few methods have shown a decrease in the number of features selected while improving classification performance measured with F1. The genetic algorithm was one of those, proving to be a robust selector even when the similarity bias is removed. Finally, an analysis of the results from all experiments is performed to find which of the statistical features are more relevant and from what sensor they come from. Standard deviation and variance from the P-MON-CKP sensor are found much frequently than the others.