DATA-DRIVEN CLUSTERING AND SUPERVISED LEARNING APPROACHES FOR DRILLING DYNAMICS FEATURES IDENTIFICATION IN DRILLING OPERATIONS

Abstract

AbstractDrilling operations in the oil and gas industry face significant challenges, including the failure of the Bottom Hole Assembly (BHA) and the wear of drilling bits, which lead to increased Non-Productive Time (NPT) and considerable costs. These challenges primarily arise from the harsh environments present in the rocks that the bit and the drill string encounter during the drilling process. To improve drilling efficiency and cost saving, a comprehensive understanding of drill bit dynamics and its governing parameters is key. Drilling vibrations are generated due to the contact between rock formations and the drill string. These are classified into three modes: axial, lateral, and torsional. This study presents an approach to classify and predict drilling vibrations through a combination of an experimental setup and machine-learning approaches modeling. To achieve this, an experimental setup was created to simulate the drilling operation, and rocks such as chalk, sandstone, and granite were drilled. The research focuses on key surface parameters that drillers can manipulate, such as revolutions per minute (RPM) and weight on bit (WOB), to mitigate downhole problems. A 2” PDC drill bit utilized in this project makes the research interesting because it allows for studying real bit behavior with actual bits and rocks. Complementing this, a gyro data sensor placed on the bit helps to record real data supported by DasyLab, a software that records information from a Data Acquisition System (DAQ), which is a group of sensors responsible for measuring and acquiring physical parameters that are later processed by the computer. The experimental setup consists of a mechanical drilling apparatus integrated with both the earlier-named gyro data sensor and DaisyLab software. This set allows for real-time monitoring and analysis of drill bit behavior under varying conditions. The test was conducted using different rock types and two ranges of parameters for RPM (above 166 and below 100) and BHA weights (69 kg and 54 kg). These values allow us to simulate drilling scenarios. Following the physical experiments, a machine-learning model has been developed to emulate and predict drill bit dynamics. This virtual representation enables exploration of a range of operational conditions and parameters beyond the limitations of physical testing. The possibility of using machine-learning facilitates the integration of historical drilling data, improving the model’s predictive capabilities. The results of this study provide valuable insights into the complex relationship between surface parameters, downhole tools, and wellbore conditions. For instance, optimizing drilling parameters, predicting and mitigating potential failures, reducing wear on drilling equipment, NPT, and costs, and enhancing overall drilling efficiency and safety.