Unsupervised Learning Algorithms Applied to Anomaly Detection in Oil and Gas Wells

Resumo

Monitoring through sensors is a powerful tool in the evaluation of vibrations, loads, deformations,
among other problems in which gathering data allows to detect undesirable events that may arise in structures.
Growing opportunities have been observed in companies offering sensing, monitoring, and digital transformation
services, which offer cost reduction, increased operational safety and improved performance. Technologies for
processing the data collected by sensors using machine learning (ML) methodologies have proven to be efficient
tools in engineering processes. In the context of petroleum engineering, the prediction and detection of unexpected
events stands out, by supporting decision-making processes and adding value to products and services. Thus,
this paper aims to study and develop ML-based models for detecting anomalous states in oil wells, by applying
classical techniques such as Support Vector Machines, Isolation Forest and Deep Neural Network. It is expected
to compare the efficiency of these methodologies applied to time series datasets of pressure, temperature and flow
rate, allowing to predict the anomaly occurrence and generate alerts to the production operator. It is observed the
practical application and potential of the proposed methodologies for the intended product, being able to improve
the fault detection process in oil wells, as well as ensure their integrity.