Open-world learning applied to oil wells using autoencoder-based clustering

Abstract

This thesis investigates the use of Open-World Learning for anomaly monitoring in oil wells, a critical aspect for operational safety and efficiency. Existing methods identify known anomalies, but handling unknown anomalies in dynamic environments remains a challenge. The objective of this work is to develop an Open-World Learning approach that integrates autoencoder-based dimensionality reduction, binary classifiers, and a hybrid clustering method to detect and organize unknown anomalies. The hypothesis is that this approach can contribute to detecting and classifying unknown anomalies in oil well production data. In the proposed strategy, anomalies are first detected through autoencoder reconstruction errors. Binary classifiers then determine whether the anomaly belongs to a known category. If it does not, a clustering method groups similar events into new categories, which are subsequently validated by human experts. Experiments with the 3W dataset and realworld oil well data indicate that the discovered clusters align with known classes, achieving 81% overall accuracy and exceeding 95% in certain cases, while updated binary classifiers reach up to 99% accuracy. The results suggest that the proposed strategy represents an advancement in monitoring and anomaly detection, thereby significantly contributing to the safety and structural integrity of oil wells.