Detecção de anomalias em poços de petróleo surgentes com stacked autoencoders
Keywords: Autoencoders, Fault detection, Oil well monitoring, Multivariate time series classification, one-class classification
AbstractThe offshore Exploration and Production (E&P) is responsible for most of the oil and gas production in Brazil. Due to the high level of complexity in this industry, new technologies have been proposed over the past few years. The present work aimed at developing systems for fault detection in offshore oil production wells. The public domain 3W dataset was used and stacked autoencoders were implemented for dimensionality reduction. Measurements of five process variables were used as inputs for classification with examples from a single class. Isolation Forest and Support Vector Machines of a class were the techniques used to detect anomalies in the process, such as hydrate in the production line. The results were compared with other works in the literature, and an improvement of up to eighteen percent was observed. Moreover, the designed autoencoders were effective in dimensionality reduction, helping to find more parsimonious classifiers.