Session: 11-03-02 Data Science Applications in Drilling

Paper Number: 125728

125728 - A Novel Approach of Data Science Incorporating Physical Knowledge for Early Stuck Detection 

The early detection of a stuck pipe, which is almost in agreement with the prediction, is challenging and crucial as it is one of the major incidents resulting in nonproductive time and in the worst scenario, well abandonment. Some researches adopted an ordinal supervised machine learning approach using datasets for “stuck” and “normal”. However, for early detection before stuck occurs, the application of the ordinal supervised machine learning has several concerns, such as limited stuck data and lack of an exact “stuck sign” which should be a label in the training dataset. Therefore, our previous studies proposed the unsupervised machine learning approach using only data on the normal activities and adopted algorithms such as multiple prediction models and autoencoder with long short-term memories. In addition, our previous study attempted to apply graph network aiming at expressing the dependencies between multiple drilling data. These pure data-driven approaches provide a possibility to predict the stuck pipe incidents, however, the prediction performance including the frequent false alarms should be improved.

This study proposes a novel approach of data science approach, including machine learning, incorporating physical knowledge. Since the measured drilling data contains uncertainties and is affected by operating conditions, the focuses drilling parameters are expressed using the physical knowledge with unknown parameters, and the parameters are determined by data science technologies from the nearby past hook load data. The physical knowledge explains universal facts, and it is expected to play a role of data macroscopic phenomena. On the other hand, data science approach is expected to provide the local feature. Our proposed approach focuses on certain drilling parameters that could express the features of stuck events, such as standpipe pressure, drilling torque, or hook load, and these drilling parameters are expressed with other drilling parameters from the physical knowledge such as the torque and drag model with unknown parameters. Here it should be noted that there are several stuck mechanisms such as pack off, differential pressure, keyseating, and so on. Each mechanism is considered to exhibit its own characteristics in the drilling parameters, and based on this, the drill parameters of interest are selected and expressed by physical model.

We have been developing the prediction models focusing on the standpipe pressure, hook load, and drilling torque by establishing the physical models to express these parameters. These models demonstrate the higher prediction performance including fewer false alarms than pure data-driven approaches.

In this study, we firstly introduce the stuck prediction by data-driven approaches, both supervised and unsupervised machine learning approaches that we have examined. Then, this paper presents the summary of data science incorporating physical knowledge, focusing on standpipe pressure, hook load, and drilling torque, and also demonstrates the early stuck detection using the filed data containing the stuck pipe events.

Presenting Author: Tomoya Inoue JAMSTEC

Presenting Author Biography: Tomoya Inoue is a senior researcher at JAMSTEC (Japan Agency for Marine-Earth Science and Technology) and a visiting professor at Kobe University. He was involved in the drill ship project as a chief engineer of drilling system as well as a full-depth ROV project as a project leader. His research interests are in offshore deep drilling technologies including machine learning and drill pipe dynamics.

Authors:

Tomoya Inoue JAMSTEC
Tatsuya Kaneko JAMSTEC
Yujin Nakagawa JAMSTEC
Ryota Wada The University of Tokyo
Shungo Abe JOGMEC
Gota Yasutake JAPEX