Detecção de anomalias em poços de produção de petróleo offshore com a utilização de autoencoders e técnicas de reconhecimento de padrões

Abstract

The Exploration and Production (E&P) offshore segment of the Oil & Gas industry is responsible for most of the oil and gas production in Brazil. Due to the high level of complexity in this industry, it has been demanding new technologies over the past few years. This work aims to develop systems for detecting and classifying failures (anomalies) in offshore oil production wells. Two real data sets are considered in this work. The first set consists of wells operated with artificial lift by gas lift. The second approach uses public domain data 3W dataset that were collected from production wells with natural elevation. Stacked autoencoders are used in artificial elevation gas lift wells to reduce dimensionality and pattern recognition techniques such as k-NN and decision tree for an unknown failure in an oil well. After the development of these classifiers, part of the recall values obtained are greater than 0.98, which shows the applicability of the proposed system in detecting flaws in non-emergent production wells. For emergent wells, stacked autoencoders were used to reduce dimensionality. The data after the treatment were used as inputs for classifiers of only one class (one-class) such as SVM and isolation forest in order to detect anomalies in the process as hydrate in the production line. The results of the F1 score averages presented by the models are compared with other works published in journals and congresses where an improvement is observed in relation to the other proposed approaches. Autoencoders were effective for problems in detecting and classifying anomalies in offshore oil production wells, presenting satisfactory results.