Gated recurrent unit-informer with mixture of experts: A novel architecture for real-time drill string vibration monitoring via surface-downhole parameter fusion

Real-time and precise monitoring of drill string vibration states is crucial for ensuring the safe and efficient drilling of complex oil and gas wells. However, limited by the high cost and poor timeliness of downhole measurement data acquisition, existing methods predominantly rely on single-source parameters, making it difficult to precisely monitor vibration modes. A novel architecture for real-time drill string vibration monitoring via surface-downhole parameter fusion was designed. First, a downhole three-axial acceleration prediction model was established based on the gated recurrent unit (GRU) to provide real-time downhole data for the model. Second, the variational modal decomposition (VMD) method is introduced to extract the multicomponent frequency domain features of downhole vibration signals. Third, a hybrid network architecture with the GRU and Informer models in parallel is designed to capture and fuse the multidimensional and multiscale features of the surface-downhole information. Finally, a Mixture of Experts (MoE) architecture is introduced to synthesize the coupled correlations between vibrations to improve the computational efficiency and accuracy of the model. The results demonstrate that GRU-Informer with MoE outperforms current models, achieving an accuracy of 0.912 and a precision of 0.820. Compared to the models with separate inputs of surface mud logging and downhole vibration data, the model's accuracy was improved by 15.59% and 11.08%, respectively. Comparative studies validate the contributions of VMD, GRU-Informer structure, and MoE, emphasizing the critical role of the designed framework in improving the model adaptability and performance. This work provides significant theoretical support for mitigating drilling risks and enhancing operational efficiency.