ROP Modeling Using Neural Network and Drill String Vibration Data

Abstract Vibrations are caused by bit and drill string interaction with formations under certain drilling conditions. They are affected by different parameters such as weight on bit, rotary speed, mud properties, BHA and bit design as well as by the mechanical properties of the formations. During the actual drilling process the bit interacts with different formation layers whereby each of those layers usually have different mechanical properties. Vibrations are also indirectly affected by the formations since weight on bit and rotary speed are usually optimized against changing formations (drilling optimization process). Therefore it can be concluded that for optimized drilling reduction of vibrations is one of the challenges. A fully automated laboratory scale drilling rig, the CDC miniRig, has been used to conduct experimental tests. A three component vibration sensor sub attached to drill string records drill string vibrations and an additional sensor system records the drilling parameters. Uniform concrete cubes with different mechanical properties were built. Those cubes as well as a homogeneous sandstone cube were drilled with different ranges of weight on bit and bit rotary speed. The mechanical properties of all cubes were measured prior to the experiments. During all experiments, drilling parameters and the vibration data were recorded. Based on analyses of the data in the time and the frequency domain, linear and non-linear models were built. For this purpose the interrelations of sandstone and concrete mechanical properties, drilling parameters and vibration data were modeled by neural networks. Application of sophisticated attribute selection methods showed that vibration data in both, time- and frequency domain, have a major impact in modeling the rate of penetration.