Automated Fracture Identification in Borehole Image Logs Using YOLO and PCA-Based Anomaly Detectione

Summary This study introduces a hybrid machine learning framework for automated fracture detection in borehole resistivity image logs, utilizing data from ten siliciclastic wells in Alberta, Canada. The methodology combines YOLO (You Only Look Once), a deep learning model, for real-time fracture localization in high-resolution resistivity images, with Principal Component Analysis (PCA) for anomaly detection in conventional petrophysical logs (e.g., GR, DT, ILD, NPOR_LIM). Resistivity logs were processed by normalizing them to 8-bit grayscale and segmenting them into 640×640 pixel frames. YOLO, optimized through transfer learning over 100 epochs, produced high-confidence bounding boxes, while PCA detected subtle fracture-related anomalies using reconstruction errors. Performance evaluation on test wells T1 and T2 yielded F1 scores exceeding 0.80 and depth RMSE values under 0.5m, demonstrating robust accuracy.