Drilling Conditions Classification Model Based on an Expert-Knowledge-Fused Convolutional Neural Network with a Multihead Attention Mechanism

Summary In this paper, we propose a drilling condition classification model based on the fusion of expert knowledge and convolutional neural networks (CNNs) with multihead self-attention (MHA) mechanisms. This model is designed to address common limitations in traditional drilling condition classification methods, including the underutilization of prior knowledge, incomplete capture of both short- and long-term time-series information, and the inherent long-tail distribution of drilling condition data. According to drilling manuals and expert experience, drilling conditions can be categorized into four primary classes of “ream/drilling,” “rotating,” “circulating/drilling,” and “trip in and out.” Based on this, initial logical constraints and sample prescreening are performed. For each preliminary category, an independent CNN with a multihead self-attention mechanism is trained to extract local physical features and model global temporal dependencies. This hierarchical design improves both the classification accuracy and robustness of the model. To evaluate the generalization ability and engineering practicality of the proposed method, comparative experiments are conducted using data from unseen wells. The results show that, without any manual threshold tuning, the model consistently achieves accuracy, precision, recall, and F1 scores exceeding 0.92. These outcomes significantly outperform three widely used data-driven baselines: CNN, CNN-long-short-term memory (LSTM), and transformer. Notably, the proposed method enhances the recall for minority drilling events, mitigating the effects of long-tail data imbalance. This indicates that combining domain-specific prior knowledge with deep feature learning can effectively reduce bias caused by data imbalance. The model also demonstrates excellent stability and real-time performance, making it suitable for online field deployment. Overall, the proposed framework offers both logical interpretability and high-precision classification capability, providing a solid theoretical foundation and engineering basis for intelligent classification and real-time monitoring of drilling conditions.