A Hybrid LSTM-XGBoost Approach to Drilling Operation Condition Recognition

Accurate recognition of drilling rig operating conditions is essential for improving efficiency and ensuring safety. Conventional methods often suffer from limited accuracy and poor adaptability under complex conditions. To address these challenges, this study proposes a novel hybrid framework that combines Long Short-Term Memory (LSTM) networks and eXtreme Gradient Boosting (XGBoost). The key novelty lies in leveraging the strengths of deep sequential modeling and gradient boosting: LSTM effectively captures temporal dependencies in drilling data, while XGBoost enhances classification accuracy and robustness. Based on information entropy, seven critical parameters - current depth, power head displacement, give pressure, give speed, rotary speed, pump pressure, and pitch angle - are selected as input features. Using real-world drilling datasets, the framework is applied to classify five representative operating states: idling, normal drilling, stuck pipe, single connection, and circulation. Model hyperparameters are tuned via grid search to ensure balanced performance. Experimental results show that the proposed hybrid approach achieves over 90% in accuracy, precision, and F1-score, demonstrating its effectiveness and reliability. This study highlights the potential of the LSTM-XGBoost framework for intelligent condition monitoring and improved safety management in coal mine drilling operations.