Supervised data-driven approach to early kick detection during drilling operation

Muojeke, Somadina Innocent (2020) Supervised data-driven approach to early kick detection during drilling operation. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (4MB)

Abstract

The margin between pore pressure and fracture gradient in new offshore discoveries continues to get narrower. This poses greater risks and higher cost of ensuring safety of lives, facilities, and the environment. The 2010 Macondo blowout has fueled increased interests in monitoring downhole parameter for early kick detection. Early detection of kick is important part of the process safety. It provides opportunity to activate safety measures. However, after an extensive literature search, certain gaps were identified in early kick detection research. This ranged from limited availability of downhole drilling data from oil fields with downhole pressure and flow measurements for research purposes to limited modelling efforts that applies machine learning to downhole measurements in the area of early kick detection. Leveraging machine learning is crucial because of the tremendous advancements in artificial intelligence and information technology. This research provides a simple design approach to build machine learning kick detection models. In the absence of field data, we collect data from existing and new experiments that records downhole measurements. A simple model is rewarding when data processing is done downhole. The hardware used is typically battery powered. Simpler and fewer software operations will lead to less power consumption, smaller memory and simpler cooling requirements. This will lead to an increase battery run time, miniaturized designs/reduced bulk size, reduced maintenance frequency for such hardware, improved response time and lower costs. In this thesis, we investigate the simplest supervised neural network-based machine learning kick detection system to ensure high reliability using experimental data. Building upon previous kick experiments conducted using a Small Drilling Simulator (SDS), we present a detailed design of a new kick experiment setup that uses a Large Drilling Simulator (LDS) and synthetic rock samples. We also provide a detailed design of synthetic rock sample with geometrical capability to trap high-pressure formation fluid within. The experiment setup produces new set of data from downhole parameter monitoring that will be used in testing the machine learning model. Parameters such as mud flow-out rate, conductivity, density, and downhole pressure from two previous drilling experiment that monitored downhole parameters are combined to build a data-driven model for early kick detection. This model combines an Artificial Neural Network (ANN) with a binary classifier at its output. Several input combinations are trained and tested. The model can be scaled to capture other types of drilling problems such as lost circulation and also applied in the LDS system. The model was tested and evaluated with data from the SDS system, SDS system with faulty conductivity data and different experimental drilling system. Abnormal pressure and flow regimes in the wellbore provide early warnings and are shown to be more significant parameters than others; however, solely relying on them can increase susceptibility to false alarm.

Actions (login required)

View Item