瞬变电磁数据携带着重要的地质信息，其资料解释和反演过程是获得地层信息和圈定异常范围的必经过程和重要技术手段。目前，广泛使用的多种一维反演方法均存在耗时长、参数难以调控、过于依赖初始模型等缺陷，因此研究出更简便、实用的反演手段是亟待解决的问题。

论文以神经网络的前沿信息为导向，以其在地球物理领域的各项应用为基础，提出了一种基于注意力机制的卷积-双向长短时记忆神经网络（AC-BiLSTM）瞬变电磁反演方法，充分利用时间差，在获得一定量数据时进行模型训练，在观测时间之余对当下采集数据进行神经网络反演，从反演耗时、解释流程等方面提高了工作效率。论文取得以下成果：

（1）在分析卷积神经网络（CNN）和长短时记忆神经网络（LSTM）发展现状的基础上，研究了神经网络的相关理论，重点阐述了其基本结构、训练方式、优化算法及相关参数，对其在瞬变电磁反演解释中的作用做出了总结。

（2）基于卷积神经网络和长短期记忆神经网络搭建了编码器-解码器模型，以二维卷积神经网络和双向长短时记忆神经网络作为编码器，长短期记忆神经网络为解码器，期间引入注意力机制（Attention）对长短期记忆神经网络隐藏层输出数据进行重点提取，最后经全连接层获得反演数据。

（3）分析瞬变电磁数据的时空特性，以测井数据为基础，构建了符合地层电性特征的正演模型并获得模拟数据集。网络搭建完成后，以监督学习方式将采样时间-视电阻率作为网络的输入信息，以Occam反演结果作为学习目标，验证了神经网络反演算法在理论层面的合理性。

（4）基于陕北研究区实测数据对论文中Occam、CNN、LSTM及AC-BiLSTM算法的反演效果进行验证，结果表明：AC-BiLSTM算法能清晰反映沉积稳定区域不同地层电性特征及含水层富水异常特征，能快速获得分辨率较高的电阻率成像结果，其各项评价指标优异，精度略高于其他三种算法，耗时远小于Occam算法。

论文设计的AC-BiLSTM反演算法为瞬变电磁数据解释提供了新的思路，一定程度上提高了瞬变电磁数据解释水平和反演效率。

Based on various applications in the field of geophysics, a convolution-bidirectional long-term memory neural network (AC-BiLSTM) transient electromagnetic inversion method based on attention mechanism is proposed, which makes full use of the time difference, conducts model training when obtaining a certain amount of data, and inverts the collected data after observation time to improve work efficiency. The following results are obtained from the research of this paper:

Guided by the frontier information of neural network and based on its various applications in the field of geophysics, this paper puts forward a convolution-bidirectional long-term memory neural network (AC-BiLSTM) transient electromagnetic inversion method based on attention mechanism, which makes full use of the time difference, conducts model training when obtaining a certain amount of data, and conducts neural network inversion on the current collected data after observation time, thus improving the work efficiency from the aspects of inversion time consumption and interpretation flow. The paper has achieved the following results:

(1) Based on the analysis of the development status of Convolutional Neural Network (CNN) and Long-term Memory Neural Network (LSTM), the related theories of neural network are studied, with emphasis on its basic structure, training mode, optimization algorithm and related parameters, and its role in transient electromagnetic inversion interpretation is summarized.

(2) Based on convolutional neural network and long-term and short-term memory neural network, an encoder-decoder model is built. Two-dimensional convolutional neural network and bidirectional long-term and short-term memory neural network are used as encoders, and long-term and short-term memory neural networks are used as decoders. During this period, Attention mechanism is introduced to extract the hidden layer output data of long-term and short-term memory neural networks, and finally the inversion data is obtained through the fully connected layer.

(3) The temporal and spatial characteristics of transient electromagnetic data are analyzed. Based on the logging data, a forward model conforming to the formation electrical characteristics is constructed and a simulation data set is obtained. After the network is built, the supervised learning method takes the sampling time-apparent resistivity as the input information of the network and the Occam inversion result as the learning goal, which verifies the rationality of the neural network inversion algorithm in theory.

(4) The inversion results of Occam, CNN, LSTM and AC-BiLSTM algorithms in this paper are verified based on the measured data in the northern Shaanxi research area. The results show that AC-BiLSTM algorithm can clearly reflect the electrical characteristics of different strata in the sedimentary stability area and the abnormal characteristics of aquifer water abundance, and can quickly obtain high-resolution resistivity imaging results. Its evaluation indexes are excellent, the accuracy is slightly higher than the other three algorithms, and the time consumption is far less than Occam algorithm.

The AC-BiLSTM inversion algorithm designed in this paper provides a new idea for transient electromagnetic data interpretation, and improves the interpretation level and inversion efficiency of transient electromagnetic data to some extent.