An enhanced data-driven framework for early kick detection based on imbalanced multivariate time series classification

Early kick detection (EKD) is viewed as an effective way to prevent blowouts in the drilling industry. Data-driven EKD methods are increasingly attracting interests from both academia and industry. However, the available kick data are usually sparse, heterogeneous, and high-dimensional, restricting the efficient application of data-driven EKD methods. To address the issue, we propose a novel EKD method named PRIL (Practical Internal Features Learning Framework), which can fully exploit the implicit feature of sparse kick data and can be deployed to diverse wellbores. Specifically, we exploit the sparse kick data in two ways: (1) adopting a robust scale method to improve the generalization performance of PRIL; (2) using a hybrid-sampling method (including over-sampling and under-sampling) to mitigate the impact of imbalanced kick data on a single wellbore and the risk of over-fitting. In addition, we integrate the domain knowledge of kick prediction into a classification model named InterLearn to further improve the EKD accuracy. Finally, our method is experimentally evaluated on a real-world dataset, and the experimental results indicate the effectiveness of PRIL and superiority of InterLearn against the conventional data-driven EKD methods.