Machine Learning Based Intelligent Downhole Drilling Optimization System Using An Electromagnetic Short Hop Bit Dynamic Measurements

Abstract

Abstract The primary objective of this research is to develop an advanced real-time advisory system to help drillers make more effective decisions and optimize Rate of Penetration (ROP), thereby improving overall drilling performance. Transformational digital technologies such as distributed processing and machine learning techniques have been utilized in developing the ‘brains' of the system. This, combined with robust electronics and high-speed short-hop electromagnetic (EM) telemetry system, enables the development of a downhole drilling optimization system. A finite element model based on the concept of ‘transfer impedance’ was implemented in MATLAB to develop the Earth/Drillstring (E/D) model, which was used for validating the short-hop module prototype before performing field tests. With high-speed telemetry and processor downhole, the system has the capability to process and analyze raw sensor signals at the bit, improving the data transmission rate, actuation delay, and response time. Successful Machine Learning models were determined by ensemble learning of multiple classifiers that are trained and tested on real-world drilling data. Four dysfunction types are targeted in this study - Whirl, Stick-slip, Axial vibrations & Lateral vibrations. The limitation of current drilling advisory systems in the industry is that they rely mostly on rig instrumentation data available at the surface due to lack of a reliable, consistent, and, more importantly, cost-effective downhole telemetry system that can offer high-speed and high bandwidth. Our main contribution to overcome this limitation results from the fact that we are processing the high-frequency data downhole at the bit utilizing advanced pre-trained machine learning models to identify drilling dysfunctions in real-time. The output of the models is transmitted using a state-of-the-art high-speed EM telemetry system, facilitating a real-time advisory system for drillers at the surface. Drilling data from several wells were used to train the models using classification algorithms - Logistic Regression, KNN, Decision Trees, Random Forest, Artificial Neural Networks, and Naive-Bayes classifiers. More than 30 parameters were used in the models with Weight-on-Bit, RPM, Torque, Mechanical Specific Energy playing a major role. Results from each of the classifiers are compared for accuracy and complexity. The highest accuracy of 89% was achieved in successfully identifying dysfunctions downhole when tested on real-world data. A library of bit dysfunction conditions and corresponding corrective actions for Whirl, Stick-slip, Axial vibrations & Lateral vibrations were developed using a hybrid of data-driven and physics-based techniques. In this paper, we introduce a novel technique to overcome two main constraints - speed and bandwidth, faced by most drilling optimization systems. A short-hop EM telemetry system ensures high-speed downhole-surface communication links, and an advanced processor with bit dysfunction libraries eliminates the need for high bandwidth. Design of the telemetry system, simulated models, and machine learning algorithms presented in the paper aid in developing a highly cost-effective intelligent drilling advisory system, which can lead to improvements in safety and non-productive time of drilling operations.