Navigating Exploration with Realtime Drill Bit Advisory Using Deep Learning AI Model

ABSTRACT This study builds on previous work of development and field trials of a Deep Learning model that predicts the health of drill bits during real-time operations using offset surface and subsurface data. We demonstrate that given limited labelled training data, the approach can be extended to bit types that are not present in the training set. We also explore the effects of the choice and quality of offset data used for training the model. In conclusion, validation of model efficiency on a real time lateral is presented. The model employs an unsupervised bi-directional Long short-term Memory-based Variational Autoencoder (biLSTM-VAE) to project surface and subsurface logs onto a lower-dimensional latent space, from which real time bit condition is deduced with the help of an XGBoost classifier. High-res time and depth series data from multiple runs are converted to a standardized form and fed into the encoder-decoder deep learning AI model which produces coordinates in an encoded 2D or 3D latent space. The choice of offsets and input features determines the extent of separation between the bit wear states on the latent space. Effects of proximity versus relevance of data with respect to coverage of formation characteristics and properties are explored for parameters like gamma, density and porosity. Change in ROP with respect to gamma/porosity is explored as well, and correlated with deep learning models’ predictions to help determine the onset of bit degradation for each run. To ensure ease of applicability, most commonly available real time surface and subsurface data are used. Despite limited labels on bit dull condition for each of the offset run (bit health can be determined only when the bit is pulled out of hole and analyzed by visual inspection), the approach is found to be robust and transferable to unseen bits. Offset and feature selection in terms of formation characteristics is found to have an influence on how the model learns the distinction between the bit condition states. Estimates - by human interpretation - of bit health downhole during the drilling process can be highly subjective and intricate, given not-very-well-defined changes in formation characteristics and drilling parameters. The highlight of this application is in the fact that it can aid in the human decision-making process during real time operations in a consistent and reliable manner, thereby improving drilling efficiency, optimizing maintenance schedules and enhanced bit trip planning.