Automatic Kick Detection Using Artificial Intelligence

Abstract Complications related to kick account for more than half of total well cost. Kick is unintended influx of formation fluids into the wellbore when the pressure found within the formation is higher than the mud hydrostatic pressure. If the event cannot be detected and controlled in time, they will lead to increased non-productive time (NPT) and costs. For example, the maximum killing pressure required will be more than two times higher if the kick is detected 15 min later. To kill a kick, actions like circulating the well with heavier/lighter mud will require the field engineers to stop drilling. Even worse, a kick can potentially result in a blow-out event, which can lead to loss of life and a company's reputation. In recent years, people have made dramatic progress in kick detection. In general, they can be divided into physical-based models and data-driven approaches. For physical-based models, predicted values of the kick indicators are calculated and compared with the actual values. However, most of the physical-based models are unable to deal with complex situations, such as directional drilling and change of formation, temperature change, compressibility of fluids, etc. The pattern recognition approaches mainly consider the flow pattern. Moreover, these methods only recognize simple patterns, which might not be enough for real field application, where the data can be dirty and noisy. The models’ performance can also be severely impacted if the labels are unbalanced. We propose to build a real-time alarm system for early kick detection using artificial intelligence and real-time drilling data. The goal is to use this automatic alarm system to predict kick events ahead of time and hence reduce NPT. First of all, the historical drilling data need to be integrated with daily drilling reports to recognize the kick events and routine daily operations. Second, calculating the possibility of kick from the drilling data, by applying our statistical method. Then, the preprocessed data is divided into training/development/test set. The Long Short-Term Memory Fully Convolutional Network (LSTM-FCN) need to be built, trained, tuned and validated with the drilling data. We have already trained and applied the method in Chang Qing oilfield. It calculates the possibility of kick events and raise an alarm when the possibility surpasses a threshold. The accuracy for this field application is 90%.