The Centre for Innovative Ultrasound Solutions, part of Trondheim-based NNTU, the Norwegian University of Science and Technology gas been working on machine learning-assisted interpretation of cement bond logs. The project is 50% funded by Equinor whose engineers also work on the software tool. In his blog, CIUS researcher Erlend Viggen reports that the tool is now in active use by Equinor’s cased hole logging group.
Initially a convolutional neural network was trained with subject matter expert interpreted log data, tagged according to an opinion-based scale from ‘good’ to ‘poor’. However it proved hard for interpreters to stay consistent, leading to ‘overfitting’ as that the machine ‘gets the wrong idea from its training data’.
Another problem is noisy data. Noise is easy for an experience interpreter to ignore but causes problems to the computer. This issue is addressed by ‘feature engineering’ i.e. looking for a set of features in the log data that are known to be predictive. At the same time, Equinor has provided a more interpretive and granular classification schema with some 30 interpreted categories.
Viggen is now also moving from complex CNNs to simpler classifiers which have proved more robust to overfitting. Further improvement came from the integration of constraints derived from domain knowledge to remove some of the noisy data (such as spikes near casing collars and centralizers). The result is a classifier that now ‘agrees’ with human interpreters 64% of the time. The Python code for the CBL classifier will be released at some time in the future along with the rest of Equinor’s interpretation library.
Another CIUS R&D project involves the development of new technology for corrosion detection in oil and gas pipelines. An experimental setup has been established at partner Sensorlink’s facilities to develop a prototype system for testing.
© Oil IT Journal - all rights reserved.