Pipe stuck is one of the major incidents faced during drilling operations. The pipe stuck leads to non-productive time, and in the worst scenario, may lead to well abandonment. Early pipe stuck detection is a very important issue. Aiming at discussing the possibility of early stuck detection applying machine learning using surface drilling data (mud logging data), we first acquired surface drilling data during the past operation of Japanese oil and gas companies. Then, we created learning data from the surface drilling data and attempted early stuck prediction using machine learning with a variety of algorithms. This paper presents an approach, probability mixture model, among algorithms.There are three important viewpoints for machine learning. (1) Since there are very few stuck events, we may not be able to prepare enough learning data of “stuck” for supervised machine learning, (2) we can prepare a label of “stuck” using daily drilling report for stuck detection. It means the detection model providing “stuck has already happened”. However, there is no exact information of “sign” prior to stuck, and we cannot prepare a label of “sign” for early stuck detection before stuck happens, (3) pipe stuck is caused by a variety of reasons such as differential pressure, key seating, pack-off, or hole collapsing.Owing to (1) and (2), supervised machine learning is not expected to provide high performance of early stuck detection. Therefore, we apply unsupervised machine learning with only normal data except data during stuck and several hours before stuck because there may include the sing of stuck. In addition, owing to (3) a single prediction model cannot be expected to provide efficient prediction performance. We therefore adapt an algorithm consisting of multiple prediction models. Each model expresses its own unique prediction characteristics. A famous technique to obtain the final prediction result from the prediction results of multiple prediction model is ensemble learning. However, as each model is expected to possess the unique prediction characteristics as mentioned above, to obtain final prediction result of higher performance, we also adapt a gating module providing the selection probability of each prediction model. The gating model is created by machine learning as same as prediction models. That is a probability mixture model.This paper presents the unsupervised machine learning approach with the probability mixture model consisting of multiple prediction models and the gating module providing the selection probability of each prediction model. This paper also summarizes the results of unsupervised machine learning for early stuck detection using actual drilling data.

Attempt of Early Stuck Detection Using Unsupervised Deep Learning With Probability Mixture Model

Paper Proceedings: https://asmedigitalcollection.asme.org/OMAE/OMAE2021/volume/85116

OMAE 2021

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.