Abstract
Logging data play an important role in resource exploration, but in actual production, the logging attributes are often distorted or missing, which affects the geological work based on logging data. In this paper, a multivariate feature fusion model is proposed to address the problem of insufficient extraction of the intrinsic correlation between logging attributes in the traditional logging curve reconstruction method. Analogous to the method of extracting texture features in images, the Gray-Level Co-occurrence Matrices (GLCM) and Gradient Co-occurrence Matrices (GCM)are used to extract some kind of co-occurrence relationship that inherently exists between attribute values and trends, the fused features are fed into the BiLSTM network as a priori features in conjunction with the original attribute features, and the missing data are reconstructed using the trained model. The purpose of using this hybrid architecture is to extract the correlations between logging attributes of human interest in a previous step of network training, and to translate this type of important information into a form that can be learnt by the neural network, helping the network to learn more directly about the effective features in the logging data. In order to verify the effectiveness of the model, this paper uses some common evaluation metrics to assess the feature extraction results and model effectiveness. The experimental results of the test boreholes show that the reconstruction results of this paper's model compared with some common machine learning models (XGBoost, RNN, LSTM, BILSTM) are improved up to 33.18% and 44.12% in accuracy (MAE), and 21.67% and 19.85% in relevance (PCC), which has a practical engineering has some application value.