Machine Learning Methods for Kick Detection
Date
2023-01-20
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Early kick detection is crucial for a safe drilling operation while alarming kick events and helping to prevent blowouts. This research uses machine learning methods to develop data based kick detection models. OpenLab simulator used to create the Kick and NoKick class training and test data; the simulator drilling parameters used: Delta flow, Hook load, Pit volume, WOB (Weight On Bit), ROP (Rate Of Penetration), RPM, SPP (Standpipe Pressure) and Torque. This research selected the following machine learning algorithms: Decision Tree, Naive Bayes, Logistic Regressions and Neural Network. MATLAB was utilized for training and testing the different machine learning kick detection models. For each of these kick detection models, performance is calculated using precision (Kick classes truly predicted divided by the sum of kick classes truly predicted and Kick classes falsely predicted) and recall (Kick class truly predicted divided by the sum of kick classes truly predicted and NoKick classes falsely predicted) metrics. This research used a modified MATLAB radbas function as a reward function for the neural network model transfer functions to improve the model performance. Also, new model performance metrics were developed to evaluate the kick detection models in a more realistic way than precision and recall. These new metrics measure the models' ability to predict the kick event with an adequate early or late kick detection threshold time. MATLAB runs showed that decision tree and neural network are the most efficient models for detecting the kicks on the test data. Neural network and decision tree models' best accuracy showed an 85% KDP ( Kick Detection Performance) versus 54% for naïve bayes and 39% for logistic regression. These models' KDP were associated with the highest NPP (NoKick Prediction Performance) for both decision tree and logistic regression at 100% and naïve bayes at 92%, while neural network NPP was only 46%. A neural network model using the reward function detected 11 kicks out of the test data total of 13 kicks using a robust 10 second kick detection threshold time; this proves that the neural network model can detect the kick event quicker than the conventional kick detection systems.
Description
Keywords
Kick detection, Machine learning, OpenLab simulator, Neural Network
Citation
Abdul-Ameer, H. (2023). Machine learning methods for kick detection (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.