Multi-Fidelity Modeling and Reliability Analysis of Off-Shore Production Wells

SUMMARY & CONCLUSIONSIn an age where oil and gas industries are booming and identified as key sectors in the economy, production quantities drastically increased worldwide. This highlighted the importance of monitoring and improving the reliability of these systems as they can have a high impact on the health and safety of the surrounding environment. Many studies have been invested in long-term monitoring of production wells both onshore and off-shore, however, there are still significant gaps when addressing the monitoring of off-shore, partially observed systems. While the risks and effects of failures in such systems have been studied and outlined, there is still more work to be done on combining heterogenous data sources to improve the understanding of partially observed systems. In order to utilize all available data for system observability and analyses, the complex relationships between the data sources needs to be quantified and leveraged to predict the impact on the system reliability or state. This work aims to utilize machine learning tools to combine available data and information to understand the state of the reliability of offshore production wells as well as isolate possible root causes of failures given partial information.Research on multi-fidelity methods is readily available to identify relationships between different information sources. However, combining multi-fidelity tools with system risk and reliability monitoring tools is a large area of research that is missing. The current work provides a nexus between a Bayesian Optimization (BO) based multi-fidelity tool for system state identification and a Dynamic Bayesian Network for continuous monitoring of the health of the system and the potential impact from subsystems. Dynamic Bayesian Networks are used to model the time-dependent nature of the state of complex systems by identifying paths of component failures and the relationship between the subcomponents. This work builds upon a Gaussian-Process based multi-fidelity BO method, MFNets, that can provide a probabilistic surrogate modeling method to integrate available data sources to be used to make better predictions on the states of the effected components or subsystems. The results show that the value of leveraging this information is invaluable to understanding complex systems with very limited data. A case study of an off-shore production well is used to illustrate the significant benefit of incorporating all available information of different fidelity levels to update the beliefs of the system.