Automated Extraction of Borehole Breakout Properties from Acoustic Televiewer (ATV) Data

Abstract

ABSTRACT: Analysis of borehole breakouts is a widely used indirect method to estimate in situ stress and assess borehole instability. The Acoustic Televiewer (ATV) is commonly used to log borehole breakouts, but noise inherent to ATV data affects the interpretation of the breakouts. ATV data are usually interpreted manually with associated low efficiency and high cost, and to overcome these deficiencies here we propose a high performance combined de-noising procedure based on the median filter and deep learning techniques. We investigate the use of a convolutional neural network (CNN) method to automate the process of breakout identification, by extracting the underlying relationship between the ATV data and the borehole breakouts. We show that the CNN method can capture breakouts with reasonable accuracy, and when compared to manual procedures the new technique offers both procedural efficiency and reduced interpretation costs. 1. INTRODUCTION Breakouts occur in borehole walls when the induced stress exceeds the local rock strength. Their presence is thus indicative of high in situ stress conditions, and it is therefore valuable to interpret and analyze them in order to assess borehole stability. Two kinds of borehole imaging devices are in wide use: optical imaging tools and acoustic imaging tools. The Acoustic Televiewer (ATV), a popular acoustic imaging tool, is the source of the data used in this work. The ATV transmits ultrasonic waves out to the borehole wall, and records the travel time and the amplitude of the echo. The locations and sizes of breakouts can be obtained from interpretation of ATV data. Currently the process of breakout interpretation is conducted manually, with associated low procedural efficiency and high costs. As a means of avoiding these shortcomings, we set out to investigate whether modern deep learning methods can provide efficient tools for automatic detection and interpretation of breakouts in ATV data. Deep learning segmentation methods are widely used in many different fields, such as identifying tumor boundaries in MRI images (Ranjbarzadeh et al., 2021), delineating infected regions or lesions of lung lobes due to COVID-19 infection in X-ray and CT images (Shi et al., 2021), and detection of cars, traffic signals and pedestrians for self-driving motor vehicles (Song et al., 2018).