Detection of unexpected events in oil wells using deep learning with Autoencoders and Local Outlier Factor

Resumo

In the oil and gas industry, anomaly investigation is of great interest, as careful data analysis plays a fundamental role in preventing production losses, environmental accidents, and workforce reduction, while reducing maintenance costs. To achieve these goals, a variety of computational techniques has been applied, primarily to enhance operational safety. These techniques involve the use of data from pressure sensors, temperature sensors, and control valves installed both in wells and on production platforms connected to these wells. This work investigates unsupervised machine learning models using density-based architectures such as Local Outlier Factor (LOF), combined with autoencoders to detect unexpected events in sensors of production/injection subsea wells. The first approach leverages data from operational wells, examining pressure and temperature sensors throughout the well structure, and applies them to autoencoders for data preprocessing, thereby reducing its dimensionality and capturing important characteristics. Subsequently, these processed data are fed as input to the LOF-based architecture analyzed in this work. In the second approach, the data is inserted directly to the analyzed machine learning model, avoiding any kind of dimensionality reduction with autoencoders, followed by comparisons with previous works on anomaly detection of wells. The experiments were conducted in real cases of wells that experienced operational failures, focusing on the anomalies identified by Vargas et al. (2019) and recorded in the 3W public database. Previous studies in the literature have highlighted that the LOF algorithm has shown satisfactory performance in identifying anomalies in production wells with gas lift and in spurious closures of DHSV (Downhole Safety Valve), when compared to other approaches, even outperforming recurrent neural networks, as observed by Nascimento et al. (2020) and Aranha et al. (2023), respectively. Therefore, it is expected that LOF will also demonstrate good performance results in detecting other instances present in the database. Furthermore, its integration with autoencoders for data dimensionality reduction is anticipated as a complementary strategy to enhance the results.