Spotting A False Alarm. Integrating Experience And Real-Time Analysis With Artificial Intelligence

Abstract

Abstract Time series data from past drilling operations are an under-utilized resource in oil-companies. Time series record processes in the well, effects of machinery on the rig, and presents an opportunity to improve upon the models underlying drilling simulations as well as alarm systems. We present a proof-of-concept for reducing false kick alarms during drilling by combining a physical model with techniques from artificial intelligence. We show that artificial intelligence can be used to learn from the experience implicit in time series data. It then corrects for limitations in the physical model, which results in a more accurate prediction of the mud flow out of the well, and subsequently fewer false kick alarms. Introduction Early detection of gas influx into a well traditionally depends on measurements of the flow out and total active volume. An alarm system might take the volume of the mud pits as input and sound an alarm if the change in volume or rate of change exceeds a given threshold. Direct measurements of the flow rate in and out of the well may also be used. This may have the advantage of an earlier detection of gas influx, but relies on accurate flow meters. Typical active volume threshold values are 5–10 bbl and 3 bbl for HPHT wells. Changes in flow may have a number of benign causes, some of which third generation alarm systems try to take into account [1]. Still, inaccurate prediction of these effects as well as measurement noise, means that false alarms happen frequently on the rig. Such false alarms may lead to non-productive down-time and be a security risk in themselves, as they may divert the attention of the drilling team and erode trust in the alarm system. The sensitivity of an early-warning system, i.e. how early it can detect a kick, depends on the threshold value. A lower value gives both greater sensitivity and more false alarms, which in practice need to be balanced against each other. False alarms are therefore a major hindrance for early detection of kicks. More advanced models of the well-rig system is one answer to this problem. Accurate calculations of thermal expansion of the mud and mud column compressibility add improvements. However, more advanced models tend to require more information about the rig and well to be entered into them. Even higher levels of precision may require knowledge of peculiarities of the machinery on the rig, or of other effects that are poorly understood at the moment. On the other hand, many of these effects will make their mark on the time series recorded during drilling. A promising strategy may therefore be to use the experience inherent in time series from previous wells to increase the precision of our models. The field of Artificial Intelligence (AI) offers a set of tools for "learning" dependencies between variables and making predictions based on past recordings. In this paper we demonstrate a simple combination of AI and traditional modeling which yields fewer false alarms than each method alone.