One-Class Classifiers for Novelties Detection in Electrical Submersible Pumps

Abstract

Detecting anomalies and fault novelties is of high interest in the industry due to the scarcity of fault examples to train classification systems. In this article two algorithms for anomaly detection, One-Class SVM and Isolation Forest, are successfully used as effective methods for detecting fault novelties in problems of electrical submersible pumps. Faults in submersible electric pumps generate an enormous cost for companies in the oil and gas sector, since the cost of stopping production to change the equipment is excessive, which makes it necessary to identify problems before implementation. Empirical evaluation shows that both one-class classifiers performed satisfactorily, obtaining macro f-measure values of approximately 0.86. For comparison purposes, a Random Forest trained in a conventional binary classification manner is tested and achieved a macro f-measure of 0.95. Results show that the proposed solutions can have practical applications in the classification of problems in electrical submersible pumps, changing the way the oil and gas industry addresses this difficulty.