Case-based reasoning system predicts twist-off in Louisiana well based on Mideast analog

ERIC VAN OORT, Shell Upstream Americas; and KEVIN BRADY, Verdande Technology

Images courtesy of Nabors Drilling.

Costly drilling problems do not occur without warning. They manifest themselves over time through recognizable symptoms and patterns. Recent testing of an automated cased-based reasoning (CBR) system to predict twist-off and stuck-pipe events has shown that these precursors can be accurately identified far in advance of the problem.

On the rig in real time, the predictions yielded by these tests would have provided sufficient warning to take corrective action and mitigate impending twist-off and stuck-pipe events. In addition, the testing also showed that drilling data acquired in one region can be used to identify and predict events in other areas, meaning that it should be possible to develop CBR routines with general applicability. The results contribute to a growing body of experience in CBR technology aimed at enhancing drilling safety and efficiency.

CASE-BASED REASONING

The tests examined a novel artificial intelligence system that uses a data library of past cases and compares them continuously to the current drilling situation in order to identify event precursors. Cases include information on well history and operator best practices. The system, called DrillEdge, achieves real-time prediction through continuous identification of events using the rig’s wellsite information transfer standard markup language (WITSML) data stream. These events are then compared with the case library.

Fig. 1. Radar view showing segmentation for trouble events, case movement and best-practice advice for mitigation.

Fig. 2. Time view showing maximum-torque events and case response before the first twist-off event.

When symptoms that could lead to defined problems are recognized, the most relevant case histories are retrieved and presented to the drilling team. This information is used by the team to better interpret evolving wellbore conditions. Problem mitigation benefits from immediate access to client best practices and lessons learned. By providing the most relevant cases and linking to relevant historical documents, the system offers background for why a particular event progression is likely to occur.

To communicate in a real-time environment, cases are indicated on a circular “radar” display, Fig. 1. As stored historical cases become relevant, they enter in the form of markers from the outside perimeter of the screen. As the correlation with past events increases, the case moves closer to the center of the screen, which represents a 100% match between current conditions and the historical trouble event—signifying a very likely reoccurrence of this event. This display clearly conveys the historical case’s relative degree of similarity to ongoing operations, which should be treated as a prompt to take appropriate action to prevent a reoccurrence of the historical trouble event. The radar plot allows for the evaluation of multiple trouble events at the same time through segmentation. Moreover, the case can be easily linked to recommended best-practice advice on how to mitigate the trouble event.

While stick-slip was initially assumed to be a good predictor of twist-off events, it was found to be less important than other indicators in the cases that were captured and tested.

TESTING PROCEDURES

Testing of the drilling software began with building a library of cases using data captured from an onshore well in Saudi Arabia that had experienced twist-off events. Once the cases were built, the drilling data was replayed to validate system response to known event precursors and ensure calibration.

These cases were then applied to data acquired while drilling a gas well in the Haynesville shale of northern Louisiana where two twist-off events had occurred. The software was first blind tested to evaluate how well the technology recognized problems in the Louisiana well, and to investigate the accuracy and timeliness of the response. The test personnel were not aware of when the twist-off event would occur. The system produced strong results in both twist-off events by correctly identifying them in advance of the actual occurrence.

Torque, stalling and stick-slip problems were identified as key parameters leading to the twist-off events, as these were considered to be the most likely indicators.

To retrieve the Middle East cases, automated agents were created to detect two new events: 1) torque maxed out at the limit of the soft torque limiter, and 2) stalled-out drillstring. The historical Louisiana data was replayed via a WITSML data stream as if it were a live drilling operation. No daily drilling reports, mud log data or final well reports were provided, to ensure that the test personnel had no information about where the twist-offs occurred or what the drilling symptoms were leading up to the events.

A point for future investigation is the role played by stick-slip conditions as a precursor to twist-off events. While stick-slip was initially assumed to be a good indicator, it was found to be less important than other indicators in the cases that were captured and tested. This finding may vary in different geological environments under different drilling conditions. Note that there were marked differences between the Middle East well and the Louisiana well: The former was a 17,000-ft vertical well drilled with oil-based mud, whereas the latter was an 18,000-ft horizontal well drilled with water-based mud. Although both were drilled in environments that would be considered “hard rock” drilling (i.e., with Young’s moduli exceeding 20,000 psi), there was no correlation between geologies drilled or between bits and bottomhole assemblies used.

Following this test, a reverse blind test was conducted on the Middle East well using updated cases created from the Louisiana well. The twist-off event on the Middle East well was also accurately predicted using cases from the Louisiana well.

THE FIRST TWIST-OFF

During the simulation, a case from the Middle East well first appeared on the radar about 3 a.m. when the system detected a series of maximum-torque events (Fig. 2) that correlated with a case labeled Case 104 in Fig. 3. A second case (Case 116) was very close to breaking the 50% similarity threshold required for it to enter the radar (note that this threshold can be varied, depending on sensitivity preferences). This case was built using maximum-torque and stick-slip data, but its response was hampered due to the absence of stick-slip data for the Louisiana well.

The cases retreated from the radar at 5 a.m. when the sequence of maximum-torque events ended, and reappeared when the maximum-torque events began again around 7 a.m. Twist-off occurred at 7:53 a.m.

Fig. 3. Radar screen showing the response of cases on the radar prior to the first twist-off event.

Fig. 4. Radar view showing response of three separate cases prior to the second twist-off event. Case C111 approaches the 75% similar “inner ring” three hours before the event.

The system clearly indicated the torque events as precursors to the actual twist-off. Had the system been in place on the actual well, mitigation efforts could have begun with the initial series of precursors as early as 3 a.m., allowing the well delivery team almost five hours to prevent the twist-off from occurring.

In addition to the Middle East cases, two cases were built based on the first twist-off from the Louisiana well. When the data was replayed, these two cases tracked to the center of the radar. The accuracy of this response was, of course, expected, and demonstrates that the highest sensitivity will be achieved if cases can be customized to local conditions. However, the mix-and-match evaluation of the Middle East and Louisiana well cases clearly shows that even unrelated cases may be very useful in the event that no local data is available (e.g., when wildcat exploration wells are drilled in new areas). It also points to the possibility of formulating truly universally applicable cases.

THE SECOND TWIST-OFF

The second twist-off event in the Louisiana well was also predicted far in advance. Five cases were built from the Middle East well using data captured at different times to produce a range of response times and trigger mechanisms.

Leading up to the event, the first case appeared two days before the actual twist-off, when a sequence of 31 maximum-torque events was detected. This ultimately resulted in three time-sensitive cases from the Middle East well appearing on the radar. The events reoccurred periodically until the bit was tripped.

After the trip, drilling resumed and proceeded smoothly until about 3 a.m. on the day of the event, when a sequence of 10 maximum-torque events was detected. This caused Case 111 to appear on the radar display, followed quickly by Cases 116 and 104, Fig. 4. By 4:04 a.m., Case 111 had approached the 75% threshold on the radar display’s inner circle, indicating a very close match to the scenario that caused the twist-off on the Middle East well. The twist-off event in the Louisiana well occurred about three hours later at 7:22 a.m.

FINDING PATTERNS

The system uses events that fire in time and depth—as well as properties such as lithology, hole size and well design—to find matches between the case library and the actual well. These patterns provide an effective way to see the progression from individual, isolated periodic events (which may not cause a case to enter the radar) to series of rapid events (which typically lead to serious consequences).

Maximum-torque events (Fig. 5) can be seen in increasing frequency and severity as drilling progressed to when the first twist-off occurred at 7:53 a.m. Although on this bit run a case did not appear and stay on the radar until 4:41 on the morning of the twist-off (about three hours before the occurrence), maxed-out torque events began firing soon after drilling resumed.

After the twist-off, drilling resumed and progressed for an entire bit run without triggering any events, demonstrating the robustness of the event-recognition algorithms in avoiding false-positive responses. The absence of false positives provides important validation for the practical applicability of the system, as frequent erroneous responses and associated alerts to the well delivery team would undermine confidence in the system.

Fig. 5. Time/depth view showing that events occur in increasing frequency and severity ahead of the twist-off. The maxed-out-torque events began firing two days before the twist-off occurrence. Increasing severity is indicated by color scale from yellow to red.

STUCK-PIPE TEST

 In addition to the tests that examined globally diverse data from similar wells, another test studied the use of generic data in predicting stuck-pipe events. Generic cases are those developed by a client and released for use with other clients. Success in these tests showed that generic cases can be credible predictors of problems on wells in other regions.

In this blind test, generic North Sea pack-off cases from the CBR library were used along with a case built from the Russian subject well, which had experienced stuck-pipe problems. When the Russian well data was replayed, a case created from that well appeared on the radar about six hours before the stuck-pipe occurrence. Three cases from the generic case library appeared just eight minutes later, showing a high degree of global transferability for this particular data set of cases. The response difference is minimal in successfully predicting a stuck-pipe event six hours in advance.

CONCLUSIONS

These tests have shown conclusively that precursors to drilling problems can be accurately identified far in advance of costly trouble events. This early detection enhances drilling safety and efficiency by providing ample time to resolve the situation before it becomes an actual problem.

The research has also shown that experience with problem wells in one geographic region can be used to predict similar drilling problems in other areas. The ability to transfer experience in either specific or generic cases provides the opportunity to improve safety and efficiency in new operations.

While these tests addressed twist-off and stuck-pipe problems, similar testing and verification will involve lost circulation and pore pressure (kick avoidance) capabilities. Also, future testing will demonstrate the CBR technology’s applicability not just to trouble events, but also to reinforce positive drilling practices, by relating real-time cases to optimum historical drilling conditions.

After testing concluded, the CBR system was put in actual use for real-time monitoring of live operations in the Middle East using the infrastructure of Shell’s real-time operations centers. The system will also be ported to North American onshore gas wells in mid-2011 when 24/7 surveillance becomes available.

THE AUTHORS

ERIC VAN OORT joined Shell Research in The Hague in 1991, after earning his PhD in chemical physics, to work on shale stability problems and drilling fluid design. He now serves as Technology Manager for onshore gas operations at Shell E&P Americas.

KEVIN BRADY serves as Vice President of Sales and Marketing for Verdande Technology. He began his career with Sperry-Sun as a field engineer before moving into research and development, then sales and marketing. He earned a BS degree in geology from Louisiana State University and an MBA in marketing from the University of Houston.