Intelligent Identification First Arrivals of Acoustic Logging Curves Using Dual Attention PhaseNet

Accurately picking the first arrivals of acoustic logging curves (e.g., P-, S- and Stoneley waves) is crucial for stratigraphic lithology characterization. The conventional interpreter-dominated first arrivals identification methods frequently lead to an interpretation uncertainty and time burden. To reduce these deficiencies, we developed a dual attention PhaseNet (DA-PhaseNet) network to intelligently identify the first arrivals of acoustic logging curves. The field data test demonstrates that the DA-PhaseNet network can dramatically improve the result accuracy and its generality compared to other PhaseNet-based methods. Specifically, the DA-PhaseNet strategy can capture both local and global features of input logging curves simultaneously, resulting in a high identification accuracy of 99.4% and 94.5% for P- and S-wave respectively. Moreover, the proposed DA-PhaseNet network dramatically improves the accuracy of first arrival identification from 72.3% to 87.6% for the noise-contaminated Stoneley wave. Furthermore, it is important to mention that the DA-PhaseNet has a maximum noise tolerance of 0 dB for P- and Stoneley waves to ensure accuracy of first arrival identification, while has a maximum noise tolerance level of 10 dB for S-wave if the result F1 score limit is set at a level of > 0.8.