Unexpected influxes and losses pose a significant risk to rig personnel, the environment, and drilling efficiency. Influxes and losses typically manifest in the circulation system (e.g., as increases or decreases in flow and mud-volume) and other rig surface measurements. Automated, real-time interpretation of these drilling parameter traces is complicated by the highly variable and transient nature of the circulation system during normal drilling operations. As a result, the most commonly deployed automated alarm systems (e.g., fixed +/- bounds) have high false alarm rates, and are sometimes treated as unreliable by rig personnel. Recent advances in machine learning enable data-driven algorithms to identify anomalous behavior in real-time data traces, but until recently, the uptake of these algorithms has been hindered by driller's lack of trust in these automated systems, and the complexity of explaining why these so-called "intelligent" algorithms do (or don't) generate alarms in any given scenario. This paper documents a novel machine-learning algorithm framework for circulation system monitoring that was designed to maintain a very low false-alarm rate and earn the driller's trust by explicitly providing expected safe operating bounds on flow-out and pit-volume, so that even during long (e.g., 24-hour) periods without alarms, the driller knows that the system is operational and trustworthy. We present performance results generated across a massive body of drilling data that illustrate the trade-offs between detection rate and false alarm rate that are inherent to any machine learning (indeed, any algorithm) approach to event detection, and show how explicit bound generation can be used to improve driller trust and acceptance.

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