A well logging curve reconstruction method based on BiLSTM neural network and Multi-Head Attention mechanism

Well logging curve data often suffers from partial loss due to borehole collapse and instrument failures, while re-logging incurs high costs. To address the insufficient accuracy of existing reconstruction methods, a well logging curve reconstruction method is proposed based on Physics-Prior Driven Hybrid Neural Network (PPD-HNN) in this paper. This method captures bidirectional dependencies in sequential data through a Bidirectional Long Short-Term Memory (BiLSTM) neural network, while enhancing the attention of model to critical features via a Multi-Head Attention (MHA) mechanism, thereby improving reconstruction accuracy. A Particle Swarm Optimization (PSO) algorithm is introduced for hyperparameter tuning, with geological constraints incorporated to ensure reconstructed results comply with the physical laws of logging curves, avoiding unreasonable data fluctuations. Orthogonal experiments and well logging curve completion and generation experiments were conducted using real well logging data from Qingcheng Oilfield. The optimal model architecture and hyperparameter settings were determined through orthogonal experiments. In the completion and generation experiments, such comparative models as Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) neural networks, BiLSTM, and eXtreme Gradient Boosting (XGBOOST) were included to validate the proposed model's accuracy. The results demonstrate that PPD-HNN performs well in capturing the nonlinear relationships among logging curves and the sequential characteristics along depth, achieving an R² improvement of approximately 19% over HNN. This method offers a novel technical approach for low-cost, high-precision well logging data restoration.