| Popis: |
Borehole acoustic image logs present meaningful information about some petrophysical properties of a reservoir. The identification of these features, in addition to several supplementary data are very important to achieve a valid characterization of the reservoir itself. Among the different patterns that image logs could show, it can be mentioned fractures and breakouts occurrences, which are very important for the experts in the field. These occurrences usually appear on borehole acoustic image logs. Its study can provide some insights about petrophysical properties of the reservoir, such as the mechanical stresses borne by the geology of the site, and this, for example, can be linked to the dynamics of the fluids retained inside the reservoir itself (i.e., hydrocarbon, gas or water). Usually, the workflow used to identify these structures requires experts to manually identify them. In this work, we propose a novel methodology for automatic detection in Borehole acoustic image logs of such structures using a single Fast Region-based Convolutional Neural Network (fast-RCNN), which could easily be incorporated into the analysis workflow carried out by petrophysicists, contributing to the characterization process. Due to the high number of samples that this type of neural network require, we trained our algorithm simulating pattern breakouts and fractures and after we test it in real and simulated images. Our model reported the area under the Receiver Operating Characteristic curve (AUC) of 98% and 90% for detecting fractures and breakouts, respectively, in simulated data. The algorithms commonly reported in the literature to identify fractures and breakouts, only achieve around 81% and 73% of AUC, respectively. Finally, for real images, the fast-RCNN detected efficiently the fractures and parts of breakouts. Thus, our data suggest that deep learning can improve significantly the accuracy on automatic detection of geological formations in acoustic borehole image logs. Available repository at [16] . |
Full Text from ScienceDirect