Deep Transformer Networks Applied to Anomaly Detection in Oil and Gas Wells

Abstract Ensuring well integrity in offshore oil and gas operations is critical for safety, environmental protection, and production efficiency. Early detection of anomalies, such as spurious Downhole Safety Valve (DHSV) closures and rapid productivity loss events, mitigates risks and economic losses. However, challenges like anomaly rarity, data imbalances, and operational complexities, including valve changes, hinder the effectiveness of traditional detection methods. This study evaluates TranAD, a deep transformer network-based model, to detect anomalies in multivariate time-series data using the 3W dataset, a realistic public offshore well repository. TranAD leverages attention-based architecture to capture intricate temporal dependencies, making it well-suited for imbalanced datasets. When trained exclusively on non-anomalous data, it employs adversarial training and predictive-reconstructive learning for robust detection. The methodology incorporates preprocessing to address missing values, frozen sensors, and imbalances, with feature selection optimizing model performance. Two training approaches are examined: well-specific models and a generalized model spanning multiple wells. The findings highlight the potential for real-time anomaly detection systems to ensure well integrity in offshore operations. By employing advanced methodologies, operators can achieve automated, continuous monitoring of offshore wells, reducing risks and enhancing operational reliability.