Classification of undesirable events in oil well operation
Evren Mert Turan, Johannes Jäschke
Abstract
Various algorithms are compared for the automatic classification of undesirable events during the operation of oil wells. The 3W database compiled by Petrobras is used to compare classifiers and some aspects of the workflow. Classification is performed during the transient phase of the event, and with the aim to help operators identify which of seven classes of unwanted events is occurring. Features are calculated based on sensor data within a time window. Feature selection is found to not significantly affect classification accuracy. Out of the examined classifiers a decision tree performs best: most events are classified at 90+% accuracy, with an Fl-score of 85%. This decision tree has similar performance to a more complex random forest developed in prior work, likely due to choices in the feature engineering.