Conventional kick detection methods mainly include monitoring pit gains, surface flow data (flow in and flow out), surface and down-hole pressure variations, and outputs from physics-based models. Kick detection times depend on a driller’s individual ability to interpret these drilling measurements, symptoms and model predictions. Furthermore, testing a novel data-driven solution in a full-scale operation may induce non-productive time, safety risks and crew fatigue adding to false alarms that inevitably occur during testing. Therefore, the development of better, faster and less human intervention-dependent kick detection on a laboratory scale system is a valuable step before full-scale testing.
We have generated a dataset containing seven typical drilling measurements and a sequence of gas kicks from experiments conducted in the laboratory scale. First, we employ data analysis tools following data pre-processing steps, data scaling, outlier detection, and natural feature selection. Next, we consider additional “engineered features” and apply different feature combinations to logistic regression with an ensemble method (boosting) for developing kick detection algorithms.
In our data analysis, ‘Delta flow’ (difference between flow in and flow out of the well) and ‘Rate of change of delta flow’ designed features, combined with logistic regression and boosting, give promising results in detecting kicks. Finally, we propose an intelligent algorithm and alarm architecture for a complete kick alarm system, which draws from both data analysis and machine learning models developed in this work.