Background Using directional boreholes, the borehole transient electromagnetic (TEM) method enables near-field excitation and the simultaneous reception of three components within a borehole. This method can effectively avoid interference from ferromagnetic materials in roadways, thereby significantly enhancing the detection accuracy and range. Therefore, this method is widely applied to the detection of underground concealed water hazards in coal mines. However, there is a conflict between the accuracy and efficiency of current inversion methods for borehole TEM data, and existing technologies are insufficient to simultaneously meet the demands for high efficiency and high precision in the detection of concealed water hazards.

Objective and Methods To address these issues, this study developed an adaptive sample construction method based on Bayesian optimization and established an artificial intelligence-based inversion framework using a long short-term memory (LSTM). Specifically, the theoretical responses of measured data were reconstructed using the method of trend surface analysis first, aiming to reduce the impacts of factors such as environment on the data and lay the foundation for the analysis of the consistency between the distributions of measured and simulated data. Then, the Bayesian optimization strategy was introduced, and the objective function for data distribution consistency was constructed using maximum mean discrepancy (MMD) and area under the curve (AUC). This objective function allows for the quantitative evaluation of the similarity between the distributions of simulated sample data and measured data, thereby contributing to the determination of the optimal resistivity range through continuous optimization. Furthermore, an algorithm for the one-dimensional (1D) forward modeling of TEM data was developed based on graphics processing unit (GPU) parallel computing. This algorithm was employed to construct a high-quality training sample database based on the optimal resistivity range. The basic framework of the inversion model incorporated the Seq2Seq architecture established using a LSTM while also integrating the attention mechanism to fully capture the temporal dynamic characteristics of electromagnetic responses. Based on 1D inversion, the quadrants of the anomalous response characteristics of the horizontal components of borehole TEM data were determined using the K-means clustering algorithm. Based on both the derived relationship between the depth coefficient and resistivity and the acquired information on toolface angles, the 3D resistivity imaging of the inversion results of borehole TEM data was achieved.

Results and Conclusions Numerical simulation experiments reveal that the ATT-LSTM inversion yielded average relative root mean square errors (RRMSEs) of 5% or less, exhibiting higher accuracy compared to the Occam's inversion and the recurrent neural network (RNN)-based inversion. Noise resistance experiments indicate that the ATT-LSTM inversion method maintained high stability and robustness under noise interference at certain levels. Physical simulation experiments were conducted using a water tank, with a copper plate utilized as a substitute for a low-resistivity anomaly. The experiment results indicate that the constructed high-quality sample database and the ATT-LSTM inversion framework jointly allow for the accurate inversion of the stratigraphic models and the characterization of the low-resistivity anomaly. The engineering practice in a coal mine in Xinjiang demonstrates that the ATT-LSTM inversion method can effectively identify underground low-resistivity anomaly zones, with the 3D imaging result highly consistent with known geological data. Overall, the results of this study will provide reliable technical support for the detection of concealed water hazards in coal mining areas, thereby effectively ensuring efficient and rapid roadway excavation.