Assuring asset integrity through improving the accuracy of leakage source identification of a permanently installed subsea leak detection system using artificial neural networks

Abstract

Environmental concerns and regulatory controls for oil and gas exploration and production activities have been increasing with the prospecting of deep-water fields and sensitive areas, such as the artic seas. To stop any incidents developing into critical events, subsea leak detection systems are required for a fast, cost-effective, and reasonable accurate method to not only detect the leakage substance (in this case methane), but also to identify its source and location. This thesis evaluates approaches to extend the capabilities of such systems deploying methane sniffers (pinpoint sensors) in locating the leakage sources by combining their sensory information with advanced data analytics. It will assess the potential role of artificial neural networks (ANNs) in improving the accuracy of leak source identification of permanently installed subsea leak detection systems.

The study reviews the advantages and disadvantages of four different modeling techniques that have been chosen to support this task: the analytical approach, the optimization approach, the probabilistic approach and the direct inverse approach. In the evaluation phase, the optimization approach, which underlies the working principles of artificial neural networks, was identified as the approach providing the highest accuracy, shortest time to run and with the lightest demands on resources in order to identify the location of methane leakage source in subsea condition. In addition, a computational fluid dynamics (CFD) module is also introduced to generate the data that is essential for the ANN training and testing process.

This thesis contains five main experiments. The first experiment provides the use of CFD to simulate different methane leakage source locations and its area of dispersion in steady state. The next two experiments are creating and training the artificial neural network architecture in order to maximize its performance. The last two experiments demonstrate the performance of ANNs using unseen data in the presence of noise-free and noisy data sets.

The overall results lead to the conclusion that the combined approach (CFD and ANN) is a promising tool for supporting pinpoint sensors used in subsea leak detection systems to increase the efficiency of identifying leakages in calm condition. Moreover this combined approach can also tolerate contaminated data up to approximately 4% of noise.