Machine Learning Helps Flag Abnormal Pressure Loss in Real Time

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 224876, “Preventing Drillstring Twistoff: Automatic Flagging of Abnormal Pressure-Loss Signatures in Real Time,” by Jonathan Cortez, Michael Yi, and Dongyoung Yoon, SPE, Intellicess, et al. The paper has not been peer reviewed. _ Abnormal pressure loss during a drilling operation signals a failure in the fluid hydraulics-circulating system. This could be the result of a washout in the drillpipe (either in the body or connection), a problem with a bottomhole assembly (BHA) component or connection, a downhole tool failure, a surface mud pump problem, or losses into the formation. This paper describes a machine-learning approach to accurately flag abnormal pressure losses and identify the root causes. Methodology for Detecting Abnormal Pressure Loss The method for detecting abnormal pressure loss builds on a previously developed probabilistic approach to detecting washouts and pump failure. Abnormal-pressure-loss belief combines pump-failure belief and washout belief. In the described approach, the physics-based modeled pump pressure is replaced with a statistically derived modeled pump pressure. This is primarily meant to reduce false and missed alerts caused by bad or missing contextual data. While the pump-failure belief (the authors refer to this as network pump-failure belief) is obtained using the network described in the previous approach, the washout belief (the authors refer to this as the combined washout belief) has two additional components [standpipe pressure (SPP) -washout belief and what the authors call stat washout belief] in addition to the washout belief (the authors refer to this as the network washout belief) from the previous approach. Network Pump-Failure/Washout Belief. The network pump-failure and network washout belief use modeled pump pressure derived statistically. This network primarily monitors the variation of SPP with flow in and flow out to distinguish between pump failure and a washout in the drillstring. Bad data from the flow-out sensor can make pump-failure belief less reliable. SPP-Washout Belief. The SPP-washout belief is used to detect downward trends and rapid drops in the measured SPP given the current drilling parameters. This feature can detect rapid pressure drops within a stand. The drilling parameters monitored for this belief are flow-in rate, rotational speed, weight on bit (WOB), and torque. This belief is not dependent on contextual data but rather only real-time electronic-drilling-recorder data. The SPP-washout belief calculation uses 20-minute pressure, torque, and WOB queues to determine if a pressure drop has occurred in those 20 minutes while accounting for parameter changes. If the flow-in rate or rig state changes, or if a downlink event occurs, the queues are reset. Stat Washout Belief. The stat washout belief compares the measured SPP with the statistical model pump pressure, which is an estimation of SPP given the current rotational speed, WOB, and flow-in rate. The statistical model pump pressure is calculated using historical SPP data taken from the well currently being drilled and is not based on data from other wells. The historical SPP values are populated into a table based on their drilling parameters. SPP values are only added to the table if drilling parameters are consistent for a period. The stat washout belief is calculated by measuring the difference between the measured SPP and the statistical model pump pressure and normalizing it using a configurable pressure-difference threshold.