Fracture identification in reservoirs using well log data by window sliding recurrent neural network

Abstract

Detecting fractures using well logs can be difficult due to the complex response of conventional logs. To address this issue, a novel method called Fracture Identification by window sliding and recurrent neural network (FIsr) is proposed. FIsr uses window sliding to generate sequence image data for training a bidirectional recurrent neural network (BiLSTM) classifier, with columns selected from both conventional and reconstructed logs. Undersampling is applied to balance the data, as the number of fracture samples is much smaller than nonfracture samples. BiLSTM extracts features from the sequence data in two directions, considering label correlations and detecting local log anomalies caused by fractures. The prediction for each sample is based on multiple overlapping sequence images to reduce uncertainties. The proposed method is validated using a dataset from carbonate reservoirs of the Asmari Formation in the Middle East, with an accuracy of 95% and recall and precision metrics exceeding 90%. A blind well test shows that FIsr can detect all fracture zones, and the distribution of fractures along the well trajectory confirms previous knowledge of the area. The study also discusses the influence of factors in FIsr.