Building a Rig State Classifier Using Supervised Machine Learning to Support Invisible Lost Time Analysis

摘要

This paper covers the development of a key component of an internal system to report invisible lost time (ILT) metrics across drilling operations. Specifically this paper covers the development of a generalizable rig state engine based on the application of supervised machine learning. The same steps used in the creation of the production rig state engine are appled here to a smaller data set to demonstrate both the tractability of the problem and the methods used to create the rig state engine in the production system. The project objective was to provide efficiency and engineering metrics in a central repository covering operated regions. The system is designed to require minimal user configuration and management and provides both historic and near real time analysis to deliver a rich resource for offset comparison and benchmarking. Identifying rig-state is at the heart of every performance and engineering analysis system. This can be thought of as a machine learning classification problem. A large supervised learning set was constructed and used to train classification models which were compared for accuracy. A key success metric was the ability to generalise the selected model across different operations. Output from the rig-state classifier was then used to derive KPI data which was presented through a web based front end. A pilot system was then developed using agile principles allowing for rapid user engagement. Testing demonstrated that the system can support all real time operations within the company simultaneously and rapidly process historic well data for offset benchmarking. The cloud-based architecture allows rapid deployment of the system to new groups significantly reducing deployment costs. The system provides a foundation for onward data science and more advanced functionality. Minimal configuration, cloud storage and processing, combining contextual data with real-time rig data, near-real-time and historic analysis capabilities, rapid deployment, low cost, high accuracy and consistent metrics are all key and proven value drivers for the system. The output data is aso a valuable resource for additional machine learning and data science projects.