Stuck pipe incident prediction integrating an enhanced attention mechanism

Drilling efficiency is critical to the safe and economic development of oil and gas resources. However, stuck pipe incidents remain a pervasive technical failure, resulting in substantial non-productive time and elevated operational costs. While conventional data-driven methods attempt to forecast future drilling parameters, they rely heavily on manual interpretation by on-site personnel to assess sticking risks. This reliance introduces inherent delays and subjective errors, severely compromising real-time operational efficiency. To overcome these limitations, this paper proposes an intelligent, end-to-end detection framework designed to automatically identify the precursor signals of stuck pipe incidents. By mapping the continuous engineering parameter space directly to the incident space, temporal data preceding an incident are formulated as precursor signals for supervised learning. To effectively capture complex temporal dynamics and latent deterioration patterns, we introduce a segmented Transformer architecture. This model partitions continuous data into local subsequences, upgrading standard point-level attention to a sequence-level attention mechanism. Furthermore, to mitigate the severe class imbalance caused by the extreme scarcity of precursor signals, the optimization objective employs Focal Loss to improve the identification of minority-class precursor samples. Evaluated on a real-world field dataset comprising 5000 samples, the proposed model achieves an overall prediction accuracy of 94.5%. The framework also shows practical utility by issuing early-warning alerts 1.5–2.0 h prior to the onset of a stuck pipe event, providing a potential solution for proactive risk mitigation.