Advancing Anomaly Detection in Oil Production Wells with TranAD: A Deep Transformer Network Approach

Resumo

The oil and gas industry is undergoing a profound transformation, leveraging cutting-edge technologies such as artificial intelligence, cloud computing, and the Internet of Things to optimize operational efficiency and safety. Within this context, ensuring the integrity of oil production wells is paramount for operational safety, environmental preservation, and minimizing production losses. Detecting unexpected events, namely anomalies such as spurious closures of Downhole Safety Valves (DHSV) and rapid productivity loss events, in well operations in a timely manner is crucial. The integration of sensor-based monitoring and computational modeling provides vital insights for identifying and mitigating such anomalies, thereby bolstering the industry's reliability and sustainability. However, the complexity of anomaly detection in oil production wells presents significant challenges. Firstly, historical data from producing oil wells tends to be highly unbalanced, with only occasional unexpected events occurring over the well's lifetime. Secondly, frequent valve change operations during the well's productive lifespan, while expected, can substantially alter pressure and temperature behavior, potentially confounding unsupervised techniques. To address these challenges, this paper proposes the evaluation of TranAD, a deep transformer network-based multivariate time-series anomaly detection model, applied on oil production wells data. TranAD utilizes attention-based sequence encoders to detect anomalies solely based on non-anomalous training data in the context of oil production wells. This study aims to assess the effectiveness of TranAD models in detecting anomalies using the 3W database, the first public repository released by Petrobras containing rare real-world undesirable events in oil wells. This dataset serves as a benchmark for the development of machine learning techniques tailored to the inherent complexities of real-world data. TranAD models trained on the 3W dataset will be compared with established benchmark techniques to validate their efficacy in this scenario. Drawing from previous case studies where TranAD demonstrated superior performance in detection and diagnosis across various domains, along with its data and time-efficient training, it is anticipated that TranAD will yield promising results in detecting anomalies in oil production wells.