Machine Learning Algorithms Applied to Detection of Well Control Events

Abstract

Abstract When a drilling rig takes a kick or experiences lost circulation it is both dangerous and expensive. The earlier these events are detected the sooner the crew can take critical corrective action, minimizing both the danger and cost associated with the event. Detecting these events early allows the crew to take corrective action early thus minimizing both the danger and the cost associated with the event. Early detection of these events requires the crew to notice subtle changes in mud volumes and flow rates on the surface. As a kick enters the well bore and begins making its way to the surface it shows up as a gain in the volume of mud at surface and also an increase in mud flow rate out of the well. Conversely lost circulation occurs when some of the drilling mud is lost down hole. This shows up as a decrease in surface mud volume and a decrease in flow rate out of the well. These increases and decreases can be subtle when compared to the normal fluctuations in the mud system during drilling operations. The mud system undergoes significant changes in volume and flow rate as connections are made, as pipe is moved in and out of the hole, as pump rates change, and even as more depth is drilled. Traditional alarm systems that trigger on simple changes in mud volume and flow rate generate a large number of false alarms. Standard mud system alarms are not effective at detecting these dangerous events. The signature of the event can be lost in the normal variance of the data. Even if a traditional alarm sounds the crew is unlikely to take it seriously due to the large number of false alarms they have encountered leading up to the event. This paper describes a system that utilizes machine learning algorithms to maintain an accurate estimate of what mud volumes and flow rates should be during all phases of the drilling process. Alarm thresholds are calculated and adapt in real time to the current rig activity. False alarms are dramatically reduced even while enforcing tight alarm bounds that enable early detection. Since the crew is left only with meaningful alarms, they are more likely to take corrective action in a timely manner.