![]()
Abstract In unconventional oil & gas fields, relative high well density and consistent well programs lead to more data and potentials to improve operation by data-driven methods. In the fields, daily reports management systems are deployed to manage daily drilling & completion operation information, where maybe millions of reports are stored. Drilling & completion activities are recorded in the daily reports, engineers always take a lot of time to analyze abnormal activities in them for operation monitoring, contractor performance evaluation and decision making, such as pipe sticking, mud loss and influx, etc. Because of the different activity coding systems between projects from different countries and data missing problems, for big or international oil companies, it’s difficult to conduct automatic and unified anomaly detection based on the database by rule-based methods.
Text classification based on machine learning and natural language processing techniques provides new solutions to automatic anomaly detection in daily reports. According to the characteristics of the drilling & completion activity recording texts, an automatic text classification method was proposed based on 460,000 operation records of 1,700 wells from 80 projects around the world. Word2vec is used to represent the texts in numbers, over-sampling and deep learning cost function optimization were used to overcome the data imbalance. Performance of several machine learning models was evaluated by ROC curve, such as naive bayes, random forest, support vector machine, long short-term memory (LSTM) and convolutional neural network (CNN), etc. and CNN was proved to be the best text classification method for daily drilling & completion report.
Finally, a 7-layer CNN model was trained to detect anomaly in daily drilling & completion reports with accuracy of 85% and was successfully deployed in an online analysis system. The text classification model helped to shorten the data analysis time and increase the technical management efficiency of the oil company.
The result proved the potential of machine learning and natural language processing in drilling & completion reports mining and laid the foundation for more intelligent solutions, such as knowledge discovering, operation problem early warning, etc.
Site News