Combining Machine-Learning and Physics-Based Models to Mitigate Stick-Slip in Real-Time

Abstract

Abstract Downhole vibrations can lead to significant drilling problems in unconventional wells, such as frequent tool failures that increase drilling costs. Detecting these vibrations during drilling is crucial for enhancing drilling practices. One of the most destructive types of downhole vibrations is torsional stick-slip, characterized by fluctuations in bit rotation speed between zero and several times the surface rotation speed. Typically, surface data is used for real-time diagnosis of this drilling dysfunction. In recent years, Artificial Intelligence (AI)-based detection algorithms have increasingly been proposed in the literature. However, a key drawback of AI methods is their limited applicability to wells in the same field with similar geological formations and bottom-hole assemblies (BHAs). This paper aims to address this limitation and explore solutions for generalizing these approaches. It does so by employing appropriate data normalization methods, transfer learning, and the inclusion of physics-based features. Using historical offset well data, this paper presents trained and tested machine learning models capable of predicting the stick-slip index (SSI) using sequences of surface measurements and physical features. This model can be deployed in real-time on drilling rigs to provide diagnostics and recommendations.