In recent years, the drilling industry has been exploring the use of sophisticated downhole sensors and more elaborate surface sensors to improve prediction and detection of trouble events such as kicks, fracture breathing and lost circulation. The use of such sensors, however, comes at a considerable additional cost and is often difficult to justify on low-cost land-based wells. On the other hand, the conventional suite of low-cost surface sensors often drift out of calibration and, without manual intervention, can provide bad data that may lead to missed or false alarms. It would be highly desirable if the conventional sensors could still be used but with improved robustness and precision for event detection. This paper reviews existing event detection methods and their shortcomings, and proposes a new method that not only addresses these shortcomings, but also improves the quality of the sensor data gathered.
This paper presents a pattern-recognition system that allows for fast analysis of sensor patterns to identify events, and a real-time sensor calibration methodology that uses a combination of a physics- based model and machine-learning techniques to improve the quality and usability of rig sensor data. Focus is on gains/kick events in particular. The calibration algorithm analyzes the quality of sensors in a mud circulation system and calibrates these sensors to yield meaningful data. This leads to a much higher accuracy in the sensed data, e.g. by mitigating flow sensor drift. The presented event detection system can then more reliably detect a kick, lost circulation or a fracture breathing event with reduced frequency of false/missed alarms.
The proposed method has been tested and implemented on actual mud circulation sensor data from the field. Using the analyzed data we demonstrate how real-time calibration can high-grade the data from a simple flow paddle sensor into much more accurate flow rate readings. We also show how sensor calibration allows for increased confidence in the values of other sensors as well, such as the pump stroke counter readings. Ultimately, the method presented in this paper allows one to extract more meaningful information from inexpensive surface sensors and make drilling trouble event predictions that are much more accurate than currently available industry methods.