Machine learning-based rock characterisation models for rotary-percussive drilling

Abstract

Vibro-impact drilling has shown huge potential of delivering better rate of penetration, improved tools lifespan and better borehole stability. However, being resonantly instigated, the technique requires a continuous and quantitative characterisation of drill-bit encountered rock materials in order to maintain optimal drilling performance. The present paper introduces a non-conventional method for downhole rock characterisation using measurable impact dynamics and machine learning algorithms. An impacting system that mimics bit-rock impact actions is employed in this present study, and various multistable responses of the system have been simulated and investigated. Features from measurable drill-bit acceleration signals were integrated with operated system parameters and machine learning methods to develop intelligent models capable of quantitatively characterising downhole rock strength. Multilayer perceptron, support vector regression and Gaussian process regression networks have been explored. Based on the performance analysis, the multilayer perceptron networks showed the highest potential for the real-time quantitative rock characterisation using considered acceleration features.

Quantum confinement detection using a coupled Schrödinger system

“…Neural networks, a crucial branch of machine learning, are considered universal nonlinear function approximators and have played a significant role in advancing contemporary technology. Several studies have already demonstrated the ability of neural networks in various domains [1][2][3][4]. For instance, researchers have utilized neural networks to forecast trajectories by applying classical physics principles, such as Hamilton's equations.…”