瞬变电磁一维反演方法存在耗时长、参数难以调控、过于依赖初始模型等缺陷。为此，提出一种基于Attention机制的卷积−双向长短时记忆神经网络(AC-BiLSTM)瞬变电磁实时反演方法，充分利用时间差，在非观测时间进行模型训练，在观测时间对当下采集数据进行实时反演。整个过程中，以实测数据加入一定比例正演数据作为数据集，以监督学习方式将采样时间−视电阻率作为输入特征，以测井约束的Occam反演结果作为学习目标，基于卷积神经网络和长短期记忆神经网络搭建编码器−解码器模型，并针对数据特性，在解码器部分加入Attention机制对隐藏层输出数据进行重点提取，最后经全连接层获得深度−电阻率数据。研究结果表明：AC-BiLSTM算法能充分挖掘数据时空特性，快速获得符合地层电性特征的电阻率成像结果，在瞬变电磁正演数据集上的预测值与正演模型拟合优度达0.898、均方根误差18.44、平均相对误差0.065，与单一长短期记忆神经网络及Occam方法相比，拟合优度分别提高了0.086、0.176，均方根误差分别减小了2.97、9.32，平均相对误差分别减小了0.012、0.068。通过对V8电法工作站实测瞬变电磁数据的AC-BiLSTM反演，快速实现了研究区地层的精准分层、圈定了煤矿采空区分布范围，获得成果与真实情况一致。研究成果突破了传统反演方法局限性，提高了瞬变电磁数据解释精度及效率。

The one-dimensional transient electromagnetic (TEM) method is time-consuming and suffers other drawbacks such as difficult parameter adjustment and heavy dependence on the initial model. Therefore, this study proposed a real-time TEM inversion method—the Attention mechanism-based convolutional neural network (CNN) - bidirectional Long Short-Term Memory (BiLSTM) (AC-BiLSTM). By fully utilizing the time difference, the AC-BiLSTM performed model training in non-observation time and the real-time inversion of the collected data in observation time. With the measured data mixed with a certain proportion of data obtained from forward modeling as the dataset, the sampling time and apparent resistivity as the input features in the form of supervised learning, and the log-constrained Occam inversion results as the learning target, the whole process of the AC-BiLSTM method is as follows: (1) the encoder-decoder model is established based on CNN and LSTM; (2) based on the data characteristics, the Attention mechanism is added to the decoder to extract the output data from the hidden layer; (3) finally, the depth-resistivity data are obtained from the fully connected layer. The study results indicate that the AC-BiLSTM algorithm can fully dig out the spatio-temporal characteristics of data and quickly yield resistivity images that meet the electrical characteristics of strata. The predicted values of the AC-BiLSTM algorithm on the TEM dataset obtained from forward modeling showed a goodness of fit of 0.898 with the forward model, with root mean squared error of 18.44 and an average relative error of 0.065. Furthermore, compared to the single LSTM neural network and the Occam method, the AC-BiLSTM algorithm showed that the goodness of fit was improved by 0.086 and 0.176, respectively, the root mean squared error was reduced by 2.97 and 9.32, and the average relative error was reduced by 0.012 and 0.068, respectively. The AC-BiLSTM inversion of measured TEM data from the V8 Receiver enabled the quick and accurate stratification of strata in the study area and the delineation of the distribution range of coal mine goaf, with the obtained results consistent with the actual situation. Research results break through the limitations of traditional inversion methods and improve the accuracy and defficiency of transient electromagnetic data interpretion.