Real-Time Bit Performance Monitoring: A Dual-Dimensional Graph Attention Network with Multivariate Time-Series Data

Abstract Efficient bit performance is key to optimizing drilling operations. Due to the concealed nature of downhole environments and the high cost of downhole measurement tools, existing research typically relies on empirical parameters or data-driven models to evaluate bit performance, assess drilling efficiency, and detect potential dysfunction. However, empirical parameters often require subjective judgment, making real-time analysis both time-consuming and labor-intensive. Data-driven models, on the other hand, struggle to capture the complex interactions between multiple parameters, and are highly susceptible to noise, making them less adaptable to the fluctuating downhole conditions. Furthermore, such models tend to produce results that are difficult to explain. This study proposes a real-time bit performance monitoring framework based on a dual-dimensional graph attention network (DD-GAT) and multivariate drilling time-series data. This workflow comprises three core modules: data preprocessing, real-time monitoring, and analysis and interpretation. During the data preprocessing stage, expert knowledge is leveraged to perform a feature augment on the input parameters, constructing a set of features closely related to drilling conditions. In the real-time monitoring stage, a feature-temporal DD-GAT is built using cleaned and feature-augmented multivariate drilling data. This network effectively captures both the inter-relationships among drilling parameters and their temporal dependencies. Additionally, by jointly optimizing prediction-based and reconstruction-based models, the framework achieves more robust representations of drilling sequences. In the analysis and interpretation phase, the two-dimensional attention mechanism provides powerful interpretability. It enables engineers to trace and analyze downhole anomalies, reducing operational risks and improving drilling efficiency. Three case studies applied in the southwestern China block and the South China Sea block demonstrate that this method effectively detects bit issues using surface sensor data, achieving high accuracy (over 95%), with precision of 0.950 and 0.927 in the bit wear and bit balling, respectively. By analyzing the dynamic interplay and evolving trends of multivariate parameters, the model can accurately identify downhole conditions such as bit severe wear, stick-slip vibrations, and bit balling. The framework not only reduces dependence on downhole tools but also provides interpretable outputs that support real-time engineering decisions, such as bit replacement timing and vibration mitigation strategies. Moreover, compared to conventional supervised machine learning methods, this self-supervised learning framework reduces reliance on labeled data and enhances the application of the model. The proposed method offers a scalable and cost-effective solution to bit anomalies, with strong potential for deployment in real-time drilling optimization systems.