Journal
IEEE ACCESS
Volume 9, Issue -, Pages 76833-76846Publisher
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/ACCESS.2021.3082557
Keywords
Fluids; Drilling machines; Mathematical model; Rocks; Deep learning; Oils; Surface cracks; Circulation losses; deep learning; drilling operations; industrial applications
Ask authors/readers for more resources
Drilling operations involve breaking rock to deepen a wellbore for oil or gas extraction, using a drilling fluid to remove rock cuttings and maintain pressure. Lost circulation incidents are a major cause of non-productive time and can lead to hazardous drilling phenomena. Machine learning algorithms have been shown to be effective in early detection of these incidents.
Drilling operations consist of breaking the rock to deepen a wellbore for oil or gas extraction. A drilling fluid, circulating from the surface through the drill pipe and from the annulus to the surface, is used to remove rock cuttings and maintain hydrostatic pressure. Drilling fluid lost circulation incidents (LCIs) are major sources of non-productive time (NPT) in drilling operations. These incidents occur due to preexisting natural fractures (vugs, caverns, etc.) and/or drilling-induced hydraulic fractures. The initiation of an LCI could lead to other hazardous drilling phenomena, such as formation influx or kick/blowout, stuck pipe incidents, among others. LCIs are typically monitored at the rig site by observing drilling fluid levels in the fluid tanks. This manual process incurs missing the occurrence or late detection of LCIs. Machine learning (ML) and deep learning (DL) classification algorithms are powerful in processing time-series data and achieving early detection of such temporal phenomena. In this study, we performed a large-scale analysis of the surface drilling and rheology data obtained from historical wells with LCIs. This analysis includes primary and secondary preprocessing steps including, aggressive sampling, feature engineering, and window normalization to derive generalizable DL models for real-time operations. Focal loss was utilized to account for data class imbalance and train robust and generalizable models. The results obtained from different ML/DL algorithms showed that one-dimensional convolutional neural network models resulted in the best performance with state-of-the-art precision, recall, and F-1 scores of 87.34%, 73.40%, and 79.77%, respectively, on unseen test drilling data.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available