Machine Learning for Detecting Stuck Pipe Incidents: Data Analytics and Models Evaluation

Abstract

The earlier a stuck pipe incident is predicted and mitigated, the higher the chance of success in freeing the pipe or avoiding severe sticking in the first place. Time is crucial in such cases as an improper reaction to a stuck pipe incident can easily make it worse. In this work, practical machine learning, classification models were developed using real-time drilling data to automatically detect stuck pipe incidents during drilling operations and communicate the observations and alerts, sufficiently ahead of time, to the rig crew for avoidance or remediation actions to be taken. The models use machine learning algorithms that feed on identified key drilling parameters to detect stuck pipe anomalies. The parameters used in building the system were selected based on published literature and historical data and reports of stuck pipe incidents and were analyzed and ranked to identify the ones of key influence on the accuracy of stuck pipe detection via a nonlinear relationship. The model exceptionally uses the robustness of data-based analysis along with the physics-based analysis. The model has shown effective detection of the signs observed by experts ahead of time and has helped with providing enhanced stuck pipe detection and risk assessment. Validating and testing the model on several cases showed promising results as anomalies on simple and complex parameters were detected before or near the actual time stuck pipe incidents were reported from the rig crew. This facilitated better understanding of the underlying physics principles and provided awareness of stuck pipe occurrence. The model improved monitoring and interpreting the drilling data streams. Beside such pipe signs, the model helped with detecting signs of other impeding problems in the downhole conditions of the wellbore, the drilling equipment, and the sensors. The model is designed to be implemented in the real-time drilling data portal to provide an alarm system for all oil and gas rigs based on the observed abnormalities. The alarm is to be populated on the real-time environment and communicated to the rig crew in a timely manner to ensure optimal results, giving them sufficient time ahead to prevent or remediate a potential stuck pipe incident.

Deep Learning and Time-Series Analysis for the Early Detection of Lost Circulation Incidents During Drilling Operations

“…In addition, wellbore collapse may occur due to reduced pressure in the annulus resulting from lost circulation, leading to drill pipe and downhole tools being buried in the hole [6]. Loss of annular pressure may result in stuck pipe incidents [7,8], which would substantially increase the complexity of mitigation operations and hence NPT. Globally, the cost of lost circulation NPT is estimated to be US$ 2-4 billion yearly [1].…”

Well Control Space Out: A Deep-Learning Approach for the Optimization of Drilling Safety Operations

“…The three main causes of drilling non-productive-time (NPT) are 1) stuck pipe incidents, 2) mud circulation losses, and 3) well influx. Stuck pipe incidents occur when downhole force(s) prevent the movement of the drillstring [23][24][25]52]. Mud circulation losses occur when the drilling mud flows from the wellbore into the formations (due to natural or induced factors) [26,27,53,54].…”

“…Moreover, the down-sampling of the data collected at the rigs represents a substantial loss of information available for the models, as the trends may be wholly lost due to the low transmission frequency of ~0.2-1 Hz. For instance, erratic torque is an important sign for detecting stuck drillstring events, however, with a frequency of 0.2Hz, it may be difficult to detect it even by sophisticated non-linear ML models [23,24,62].…”

Early sign detection for the stuck pipe scenarios using unsupervised deep learning

“…Jahanbakhshi et al (2012) developed a support vector machine (SVM) based approach for predicting a differential stuck. Also, Heinze et al (2012), Zhu et al (2019) and Alshaikh et al (2019) developed data-driven supervised models to detect stuck incident based on Decision Tree (DT), Artificial Neural Network (ANN), and Support Vector Machine (SVM) using the surface drilling parameters. Tripathi et al (2020) presented a hierarchical approach that consists of Fuzzy Rule Based and Random Forest classifier to identify drilling activities.…”