For the real-time characterisation of an inhomogeneous impact inhibiting constraint such as downhole rock layers, an unconventional method using machine learning (ML) and drill-bit vibrations is investigated. An impact oscillator with one-sided elastic constraint is employed in modelling the bit-rock impact actions. Measurable drill-bit dynamics, such as acceleration, were acquired and processed into features and 2D-images that were later used in developing ML models capable of predicting the stiffness of impacted rock constraint. Explored ML networks include Multilayer Perceptron (MLP), Convolutional Neural Network and Long Short-Term Memory Network. Both simulation and experimental studies have been presented to validate the proposed method while using coefficient of determination (R2) and normalised mean absolute error (NMAE) as the performance metrics of the ML models. Results showed that the feature-based models had better performances for both simulation and experiment compared to the raw signal and 2D-image based models. Aside being simple and computationally less expensive, the feature-based MLP models outperformed other models having R2 values > 0.7 and NMAE values < 0.2 for both simulation and experiment, thus presenting them as the preferred ML model for dynamic downhole rock characterisation. In general, this study presents a new modality to achieving logging-while-drilling during deep-hole drilling operations such as carried out in hydrocarbon, mineral and geothermal exploration.

中文翻译：

---

振动冲击钻井系统井下岩石表征的数据驱动动态方法


为了实时表征非均匀冲击抑制约束（例如井下岩层），研究了一种使用机器学习 (ML) 和钻头振动的非常规方法。采用具有单侧弹性约束的冲击振荡器来模拟钻头冲击作用。获取可测量的钻头动力学（例如加速度）并将其处理为特征和 2D 图像，随后用于开发能够预测受影响岩石约束刚度的 ML 模型。探索的 ML 网络包括多层感知器 (MLP)、卷积神经网络和长短期记忆网络。仿真和实验研究均已提出，以验证所提出的方法，同时使用确定系数（R2）和归一化平均绝对误差（NMAE）作为机器学习模型的性能指标。结果表明，与原始信号和基于二维图像的模型相比，基于特征的模型在仿真和实验方面都具有更好的性能。除了简单且计算成本较低之外，基于特征的 MLP 模型在模拟和实验方面均优于 R2 值 > 0.7 和 NMAE 值 < 0.2 的其他模型，因此将它们作为动态井下岩石表征的首选 ML 模型。总的来说，这项研究提出了一种在深孔钻探作业（例如在碳氢化合物、矿物和地热勘探中进行）中实现随钻测井的新模式。