Leveraging Targeted Machine Learning for Early Warning and Prevention of Stuck Pipe, Tight Holes, Pack Offs, Hole Cleaning Issues and Other Potential Drilling Hazards

Abstract

Abstract Stuck pipe and other related drilling hazards are major causes of non-productive time while drilling. Being able to spot early indications of potential drilling risks manually by analyzing drilling parameters in real-time has been a significant challenge for engineers. However, this task can be successfully executed by modern data analytics tools based on machine learning (ML) technologies. The objective of the presented study is to prove and demonstrate the ability of such machine learning algorithms to process and analyze simultaneously a variety of surface drilling data in real-time in order to: a) detect anomalies, that are in most cases invisible to a human eye; and b) provide early warnings of possible upcoming drilling risks with sufficient time in advance, so that the rig crew can execute the appropriate mitigation actions. The algorithms developed have favorable characteristics, such as adaptiveness to real-time data and agnosticism to well types, BHAs, mud types, lithologies or any other specific well characteristics. This supports out-of-the-box usage, which enables scalability to large numbers of wells. Targeted sub-systems detect the current operation type (tripping, drilling, reaming, etc), and detect symptoms related to differential sticking, hole cleaning issues, mechanical sticking, pack offs, tight holes, obstructions and other risks by analyzing standard surface drilling time logs in real-time, such as hookload, WOB, RPM, bit depth, mud pressure, etc. The ML models and wider risk detection system have been demonstrated to generalize to new wells, and consistently produce high performance across those tested, without any need to pre-train the models on historical data from offset wells. The system connects to WITSML data stores and outputs warnings with specific information regarding the identified symptom of the potential drilling incident, leaving it up to the rig crew or drilling supervisor to decide how to act on those warnings. The system provides drilling engineers with live warnings on average 1.5-4 hours prior to incidents, giving rig crews enough time to react. This also allows drilling engineers to know in advance a specific source of potential risk, which assists in selecting the right strategy for implementing corrective actions. The technology's performance was successfully verified in live operations and post-drill studies on historical data on over 300 wells worldwide during the past 2.5 years, with mean recall and precision metrics of 0.986 ± 0.050 and 0.712 ± 0.181 respectively across historical test wells, and significantly reduced occurrence rates of stuck pipe incidents in both onshore and offshore operations. Real case studies for onshore, offshore, conventional and unconventional assets will be presented and discussed.