Page 29 - Machine Learning for Subsurface Characterization

14 Machine learning for subsurface characterization

4 Comparative study of unsupervised outlier detection
methods on well logs

Fig. 1.3A presents the workflow for comparative study of unsupervised ODTs.
As earlier stated, unsupervised ODTs are used in different industries, including
the oil and gas industry. Osborne et al. [12] and Ferdowsi et al. [13] showed the
importance of outlier detection for robust data analysis and data-driven predic-
tive modeling. Orr et al. [14] suggested that outliers need not be removed but
analyzed to better understand the dataset. Although there is some disagreement
on how outliers should be handled, the consensus is that outlier detection is an
important process before any form of data analytics (predictive modeling, clus-
tering, ANOVA, etc.). Performance of unsupervised ODT depends on the
choice of hyperparameters and the inherent properties of the dataset, which
is governed by the process/phenomenon that generated it. Consequently, there
is no universally best-performing unsupervised outlier detection technique.
Taking this into consideration, we conducted a comparative study of the per-
formances of four popular unsupervised ODTs on well-log data by following

FIG. 1.3 Workflows for (A) comparative study and (B) actual deployment of unsupervised outlier
detection methods on well logs and subsurface measurements.