A Deep-Learning Approach for Borehole Image Interpretation
- Kinjal Dhar Gupta (Schlumberger) | Valentina Vallega (Schlumberger) | Hiren Maniar (Schlumberger) | Philippe Marza (Schlumberger) | Hui Xie (Schlumberger) | Koji Ito (Schlumberger) | Aria Abubakar (Schlumberger)
- Document ID
- Society of Petrophysicists and Well-Log Analysts
- SPWLA 60th Annual Logging Symposium, 15-19 June, The Woodlands, Texas, USA
- Publication Date
- Document Type
- Conference Paper
- 2019. held jointly by the Society of Petrophysicists and Well Log Analysts (SPWLA) and the submitting authors
- 18 in the last 30 days
- 465 since 2007
- Show more detail
Borehole image interpretation aims to evaluate the dip magnitude and azimuth direction of geological features detected along the well and to classify them. Manual interpretation of borehole images can be time consuming and is sometimes affected by inconsistencies stemming from the different interpretation approaches used. Thus, there is a need for a robust automatic or semiautomatic approach to reduce the manual labor and increase efficiency and consistency.
Recently, deep neural networks (DNN) have been successfully deployed to address a variety of challenging problems in several fields, including computer vision. We have developed a DNN-based supervised-learning technique for borehole image interpretation to automatically detect and classify geological features from borehole images acquired from a fullbore formation microimager (in water-based mud). This interpretation workflow based on machine learning (ML) requires the geologist to label a small section of features in the well for ML training; following this, the trained machine provides labeling for the remaining depths of the well. Such machines may also be trained on previously interpreted wells with similar characteristics.
Our preliminary results show that deep-learning models can accurately detect a diverse and complex set of image attributes without manual intervention. Certain geological features can occur sparsely, challenging a direct supervised-learning approach. To address this challenge, we discuss the utility of employing a semiautomatic learning paradigm.
One of the key components of reservoir characterization is the estimation of the dip magnitude and azimuth orientation of various geological features surrounding the well. The fundamental steps in the estimation process are the acquisition of well logs by specialized tools and generation of the borehole images. In conventional dippicking, a geologist visually examines these images and manually marks the location of the geological features on them using application software tools. The dip and azimuth are automatically calculated by the application given the markings. However, this process is tedious and requires a significant amount of time and effort for the geologist. In many cases, it is difficult to identify the features precisely due to the limitations of the software. Also, there can be a band of uncertainty of individual interpretations among geologists. Thus, there is a need for a robust automatic approach that can mitigate these limitations for dip picking.
|File Size||2 MB||Number of Pages||10|