Fast and accurate kick detection during drilling operations is critical to ensure drilling safety and reduce non-productive time. Over the years, the industry has taken various approaches to address this problem. However, due to the complexity of the influx process, the problem of slow detection speed and high false detection rate still exists. While many recent works of literature have attempted to solve the influx detection problem with machine learning algorithms, only a few of them have considered the time series information in real-time drilling data. Since there may be lags of unknown duration between different drilling parameters, a properly designed time series analysis model may be able to capture their relationships and make reasonable predictions. Recurrent Neural Network with long short-term memory (RNN-LSTM) architecture is a deep learning algorithm capable of making predictions based on historical time series data. Previous studies have shown that the RNN-LSTM algorithms can be applied to real-time drilling data to reasonably predict the trends of a segment of drilling data such as the total mud pit volume. In this paper, several sensitive influx indicators are separately predicted by completely independent RNN-LSTM models based on different sets of real-time drilling parameters. These models run as ensemble learning models to continuously predict influx indicators. Then, the prediction results will be quantified, and the probability of kicks will be calculated based on the different weights for each indicator. The proposed model is tested on field data in parallel with some common kick detection models and the performance is analyzed. It is concluded that the proposed model can perform accurate influx detection and outperform some common methods in the industry in terms of detection speed.

You can access this article if you purchase or spend a download.