TEM real-time inversion based on long-short term memory network
FAN Tao,XUE GuoQiang,LI Ping,YAN Bin,BAO Liang,SONG JinQiu,REN Xiao,LI ZeLin
DOI: https://doi.org/10.6038/cjg2022P0572
2022-01-01
Chinese Journal of Geophysics
Abstract:1D transient electromagnetic (TEM) inversion method relies heavily on initial model, which lead to unclear boundaries, of anomaly body, and the inversion speed is difficult to reach the real-time level. Hence, the long-short term network (LSTM) of TEM real-time inversion method based on deep learning has been proposed. The inversion can be carried out during non-observational time periods, while real-time fine imaging can be finished during the observation time period. Taking the massive sampling time-vs resistivity data as the input file, the Encoder-Decoder model in Seq2seq model is adopted, according to the characteristics of transient electromagnetic inversion, the structure of decoder is adaptively changed, and Bandanau Attention mechanism is added to highlight the role of key information. At last, the output data of depth vs resistivity produced. We applied the inversion network to tens of thousands of three layers and five layers geoelectric model which generated randomly, the test group's three measure standard deviation are all less than 10%, the reliability of the algorithm in this paper was validated, on this basis, 2 groups of near-actual model containing local abnormal body were built, the inversion of network further used in 3D numerical simulation data. The imaging results reflecting the abnormal body boundary clearly, and the computational velocity is less than 1 s.