LSTM-Based Real-Time Drill String Stuck Prediction

Summary Stuck pipe is one of the most serious drilling challenges, resulting in significant non-productive time (NPT), higher costs, and even well abandonment. Thus, developing a reliable real-time stuck pipe prediction system is critical. This study develops an LSTM-based model to predict stuck pipes in real time during drilling operations. Unlike previous research, which relied on non-real-time accessible features such as drilling mud rheology and survey data, this work uses Mud Logging data, which is readily available online in the majority of wells, ensuring practical application. Furthermore, many existing studies use inappropriate metrics to assess model performance, such as accuracy, F1-score, and recall, which can lead to misleading results in highly imbalanced datasets. To address this, the current study employs the Area Under the ROC Curve (AUC) as its primary evaluation metric, yielding a more reliable assessment. The study systematically investigates how key model parameters, such as temporal window size and resampling techniques, affect prediction performance. The dataset was collected from a well that includes four stuck pipes, with Mud Logging data sampled at 0.5 Hz. Expert recommendations led to the selection of hook load, torque, standpipe pressure, surface RPM, and weight on bit sensor readings, all of which have a strong correlation with stuck pipes. The results show that an LSTM model with a 10-second temporal window and an undersampling strategy performs the best, with an AUC of 0.868. This value is higher than those obtained using conventional models such as support vector machines (SVM) or random forests (RF) and resampling techniques. The proposed approach effectively captures temporal dependencies in drilling data, allowing for the detection of early warning patterns associated with stuck pipe incidents. Furthermore, this study broadens the scope of stuck pipe prediction beyond drilling operations by including examples from tripping and reaming processes. A case study using real-world field data validates the model's ability to provide early stuck pipe warnings, lowering operational risks and downtime costs. The results show that a Level 2 warning was triggered 33 minutes before the stuck pipe event, followed by a Level 3 alert just 3 minutes prior to the incident. These findings demonstrate LSTM's potential for real-time stuck pipe prediction, improving drilling safety and efficiency.