Deep learning driven reconstruction of acoustic logging signal in energy exploration and development

In oil and gas exploration and development, logging curves are the key data for obtaining underground geological information. However, in actual acquisition processes, problems such as drilling fluid invasion and wellbore collapse often lead to the absence or distortion of logging data, thereby affecting their subsequent analysis and application. Well logging curves exhibit clear context-dependent characteristics. Traditional reconstruction methods are mostly based on the assumption of independent and identically distributed data and are difficult to capture the temporal dependencies between data, resulting in limited accuracy of time series modeling. Therefore, for the shale reservoir in a certain basin in the northeastern region, this paper introduces a method that combines variational mode decomposition (VMD), convolutional neural network (CNN), and bidirectional long short-term memory neural network (BiLSTM) to achieve high-precision reconstruction of logging acoustic wave signals (DT). The VMD method decomposes the logging curves into different mode functions (IMF), achieving the extraction of features at different scales; the CNN method is used to extract local features such as local morphology and change trends of IMF, obtaining high-level feature representations; the BiLSTM is used to extract the bidirectional long-term dependencies of features. By standardizing the logging data, to avoid the subjectivity of manually selecting the input logging curves, the XGBoost-SHAP method is introduced to optimize the logging curves, and an DT-targeted gradient boosting regression model is constructed using XGBoost, and the SHAP values are used to conduct game theory-based contribution analysis for each input feature, obtaining the feature ranking based on the cumulative SHAP contribution. Finally, three sensitive curves, CNL, GR, and RS, are selected as input features to construct the VMD-CNN-BiLSTM prediction model, which is applied to two test wells, achieving a fitting goodness (R 2 ) of 0.71 and 0.88 respectively. Further comparative experiments have shown that the VMD-CNN-BiLSTM model has significantly improved performance in terms of MSE, MAE, MAPE, R 2 , etc., compared to the SVR, random forest, and LSTM methods. The MSE has increased by 20.5–33.9, MAE by 1.5–2.1, MAPE by 1.6%–2.3%, and R 2 by 0.21–0.36.