Visual Analysis of Anomalous Behavior Patterns in Oil Wells Using Dimensionality Reduction with t-SNE Projection

Abstract

Listen

Translate article

Anomaly detection in production processes, especially in the oil industry, is of paramount importance due to the complexity and criticality of these systems. Producing oil wells are subject to a series of dynamic and interdependent variables, which can be monitored using sensors installed along the well. Precise analysis of these sensor records is essential for early identification of possible anomalies, as any deviation from the normal pattern can result in serious consequences, from operational failures to environmental and safety risks. Given the multidimensionality and variability of this data, the use of machine learning techniques becomes indispensable for analyzing and classifying the well's situation. However, even for computational software, anomaly identification remains a significant challenge, given the oscillatory behavior and scale disparity of the variables analyzed throughout the well's productive life. To mitigate these challenges, dimensional reduction techniques can be implemented, as they can bring some advantages, such as reducing computational cost, noise reduction, and improving classifier accuracy. A widely used tool for dimensional reduction, especially for visualization, is Distributed Stochastic Neighbor Embedding (t-SNE), which seeks to project high-dimensional data into a low-dimensional space preserving clusters, bringing similar data together by affinity through a t-distribution. For this approach, real records of pressure and temperature sensors from oil wells are used. These sensors are located at eight distinct points in the well, totaling eight dimensions that will be reduced to a two-dimensional plane. Therefore, this study aims to identify behavior patterns in multivariate data from producing oil wells visually using the t-SNE dimensionality reduction method. It is expected that through visualization, it will be possible to identify behavior patterns before and after the occurrence of anomalies and enable the development of appropriate artificial intelligence algorithms for this anomaly detection problem.