Advanced Realtime Early Kick and Loss Detection (EKLD) Through a Hybrid Data-Driven and Physics-Constrained Approach

Abstract In real-world drilling operations, rapid and precise detection of early signs of kicks and losses is crucial for safe and cost-effective procedures. Failure to promptly monitor and address kick and loss events can result in severe consequences, including drilling downtime, asset losses (including facilities and personnel), significant increase of drilling costs, etc. In this paper, we presented a novel, advanced method employing a hybrid data-driven and physics-constrained approach, with a specific focus on early kick and loss detection (EKLD) for real-time application. With the latest real-time technologies, simultaneous monitoring of various downhole and surface drilling parameters has become more efficient and convenient. However, the increasing data volume generated by complex systems with various sensing devices can often become overwhelming for the monitoring system. The constant alerts from alarm system can lead to user fatigue and, consequently, undermine the performance and reliability of the active monitoring system in practical use, as users become less responsive due to frequent false alarms. Although recent studies have focused on detecting kicks or losses using data from multiple sources collected by the rig, a gap persists between academic models and real-world applications. The hybrid model presented in this paper, utilizing the robustness of data-driven approaches and the efficiency of physics-based models, while also addressing field issues such as low data quality and inaccurate measurements. The model was tested with real-time drilling data, and field results demonstrated its strong performance in predicting both kick and lost circulation events during drilling operations from real-time data stream from the rig. A performance comparison was conducted between the traditional kick detection model and the presented model. The comparison reveals that the new model significantly reduces the false alarm rate and exhibits greater reliability than traditional models.