A Deep Learning Object Detection Method for Fracture Identification Using Conventional Well Logs

Reservoir characterization struggles with identifying fractures, a typical imbalance classification problem. To handle this issue, a novel approach called fracture identification by sliding windows and you only look once (YOLO) (FISY) is proposed. FISY utilizes YOLO, a fast and accurate object detection neural network, to transform fracture identification into an abnormality detection task. Conventional well logs from specific interval formations are used as the detected image, and adjacent depth intervals enhance the analysis. Image samples are generated via a sliding window approach. YOLO detects the minority abnormality class (fractures), addressing data imbalance issues. Each depth sample undergoes multiple scans, and the probability superposition of the sliding window approach reduces forecast uncertainty. To assess FISY, a dataset from carbonate reservoirs in the Asmari Formation, Middle East, was used. The mean average precision (mAP) evaluation metric, combining precision and recall, assesses FISY’s performance in object detection. Among five-scaled FISY, XLarge demonstrates the highest mAP at 99.16%, with a rapid prediction time of 4 ms/image. In blind well tests, FISY surpasses fracture identification by sliding window (FIS) + Transformer (TF) (TF replacing YOLO in the FISY), synthetic minority oversampling technique (SMOTE) + support vector machine (SVM), and SMOTE + random forests (RFs) by over 13% in F1 score, showcasing its efficacy in fracture identification. To further assess FISY’s generalization across various lithologies, an experiment with a clastic reservoir dataset in the Ordos Basin in China reveals FISY’s good performance (mAP 87.5%).