Machine Learning for Smarter Well Integrity Management

Well integrity monitoring, especially annulus pressure surveillance, is essential for detecting abnormalities and maintaining safe production. Modern sensors enable real-time measurement of pressure, temperature, and other parameters in the well and at the wellhead. Even so, human intervention is currently required to interpret data trends, flag issues and make decisions. Machine learning may strengthen this process by automating activity recognition and detecting anomalies in near real time, reducing reliance on manual analysis on research datasets, subject to further validation before any operational use. We compare two machine learning approaches: an Inference System and Long Short-Term Memory (LSTM), by applying them to time series data for production activity recognition and abnormality detection. The dataset, consisting of minute-by-minute measurements of pressure, temperature, and valve statuses from four oil production wells in the Gulf of America, was pre-processed through cleaning and imputation before being used for model training and evaluation. The inference system combines expert knowledge with statistically derived insights to create rules as part of a rule-based method. In contrast, the LSTM model learned from labeled input data. Both approaches first classified production activities into categories such as well shut-in-steady, well shut-in-transient, production-steady, production-ramp-up, production-ramp-down, production-tubing-side or production-annulus-side valve status change, and other well activities. This classification helped better interpret real-time well activities and increase the accuracy of anomaly detection. Pressure and temperature abnormalities primarily detected during steady-state activities, such as well shut-in-steady and production-steady, were used to assess well integrity. Here, the rule-based method relied on statistical analysis to define rules for abnormality detection, while the LSTM model achieved the same task by learning from supervised data. On the evaluated research dataset, results show that the LSTM model outperformed the inference system, achieving an F1 score of 94% for well activity classification and 91% for abnormality detection. These results reflect performance on a limited dataset and do not imply predictive accuracy in field operations. The machine-learning methods and results presented in this paper are research-oriented and do not represent the operator´s engineering standards and operating procedures. The methods detailed in this paper illustrate how automatic well activity recognition may be approached, enhances abnormality detection, supports earlier well integrity issue identification, minimizes dependence on human judgment, and supports root-cause identification of well integrity issues when combined with other techniques. Such automated routines suggest feasibility as future digital tools for a well intervention operation may play a role in future research and system development.