Prevention of Stuck Pipe Events and Robust Real-Time Identification of Root Cause Using Physics Based Models in Combination with Bayesian Network Models

Abstract

Abstract When stuck pipe incidents happen, they can drastically increase the cost of a well. Although much progress has been made in stuck pipe prevention it has not been eliminated. Therefore, there is an ever increasing need to be able to automatically estimate the potential for a stuck pipe event so that preventative measures can be taken. Additionally, when the pipe does get stuck, if the reason can be determined quickly, non-productive time (NPT) can be substantially reduced. The objective here is to combine physics-based models with Bayesian networks to better diagnose drilling events and prevent stuck pipe NPT. The approach focuses on identifying three main symptoms: pack off, overpull and off-bottom rotational drag. These are detected through three separate ML models, which in this case are primarily Bayesian belief networks. Inputs to these ML models require the use of a real-time physics-based torque and drag model as well as a real-time hydraulics model. These models require contextual data which is automatically routed to the rig platform. The features used in the ML models were arrived at after a careful detailed study of a large dataset of previous stuck pipe incidents. The dataset used for building the models were from wells in North American and Middle East land operations, and North Sea offshore operations. This helped build a generalized model which is expected to be applicable irrespective of the geographic location of the well. The contextual data that is required consists of the survey data, the bottom hole assembly information, mud properties, and casing information. While it is uncommon for such data to be generally available at the rig site for real-time analysis, the team deployed novel automated data transfer mechanisms to enable real-time data transfer in all three operation locations. Tracking the data continuously also provided a means to determine when a stuck pipe incident occurred in real-time, and using data up until the current time, the algorithm probabilistically estimated the cause. This can allow for quick remedial action, which in turn leads to reduced NPT. The approach has now been deployed in the field after validations on historical datasets. This is the first time a combination ML and physics-based approach to stuck pipe detection has been deployed at a rig site. The models are not black boxes and provide prescriptive recommendations on measures to take to avoid stuck pipes. If a pipe does get stuck, the user is provided with immediate notification of likely root cause. The paper provides details on the models used and is expected to significantly improve stuck pipe prevention in the industry.