瞬变电磁长短时记忆网络深度学习实时反演方法

TEM real-time inversion based on long-short term memory network

瞬变电磁一维反演方法对初始模型依赖大,对异常体边界反映不清晰,计算速度也难以达到实时化水平.为此,本文开展基于深度学习的瞬变电磁实时反演方法研究,提出在非观测时间段进行反演训练,而在观测时间段进行实时精细成像的瞬变电磁长短时记忆网络反演策略.以正演模拟获得的海量采样时间-视电阻率为输入数据,基于长短时记忆网络构造Seq2seq编码器-解码器模型,并针对瞬变电磁反演的问题特性,对decoder的结构进行适应性更改,同时加入Bahdanau Attention机制突出重点信息作用,获得深度-电阻率输出数据.将该反演网络应用于随机生成的数万组以上三层和五层地电模型,测试组三大衡量指标标准差均小于10%,验证了本文算法的可靠性,在此基础上,构建了接近实际的两组含局部异常体模型,将该反演网络进一步用于三维数值模拟数据,取得了对异常体边界反映清晰的成像结果,且计算速度均小于1 s.

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 Bahdanau Attention mechanism is added to highlight the role of key information.At last,the output data of depth vs resistivity is 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.