Deep Learning Real-Time Bit-Wear Model Approves to be Robust and Transferable in Hard Drilling Applications

Abstract

Abstract In a previous study, we reported the development of a bit-agnostic AI deep learning wear model and showed the successful real-time deployments in multiple field bit runs. To test AI/DL model robustness and transformability, we extended the AI/DL model to another similar hard drilling application with different hole sizes, BHAs, bit designs, and lithology. The model robustness is successfully tested using a smaller dataset of subsurface logging and surface drilling data. AI/DL bit wear model employs an unsupervised bi-directional Long short-term Memory-based Variational Autoencoder (biLSTM-VAE) to project raw drilling and formation logging data into a lower-dimensional latent space. XGboost classifier is used to predict real-time bit-wear in the pre-trained latent space. Physics-based input feature engineering is conducted to extract meaningful information from the raw data. To test the model robustness when transferring to a different drilling application, only Gamma logging data and a smaller offset training dataset is applied. The deep neural network was trained in an unsupervised manner and the bit-wear estimation is an end-to-end process. AI/DL wear model was trained on twenty offset bit runs. To test the model robustness, different hole sizes and BHAs were selected in addition to bit design and lithology changes. Unsupervised training results show biLSTM-VAE model successfully extracted bit wear features and clustered bit states in the latent space. Bit wear estimation using XGboost matches with the field results for the test dataset. In the previous real-time study, Gamma ray, density and porosity loggings were used in AI/DL model for downhole formation interpretation. The model robustness test in this study demonstrates the AI/DL model can be applied to applications with only Gamma data. BHA feature testing suggests separate AI/DL models need to be trained for RSS and motorized RSS systems. AI model robustness and transformability is crucial for the success of AI model field deployments. The model robustness test in this study shows encoding drilling data into a lower dimensional space through unsupervised learnings could improve the model robustness and overcome offset training data quality issue. AI/DL results suggest exploratory data analysis of input drilling features such as BHAs and formation loggings is important for the model accuracy.