Dual Spatio-Temporal Contrastive Learning Network With Adaptive Threshold Generation for Anomaly Detection of Electric Submersible Pump

Abstract

To improve the electric submersible pump (ESP) system’s anomaly monitoring performance, this article proposes a novel approach known as the dual spatio-temporal contrastive learning network with adaptive threshold generation (DSTCL-ATG). Unlike previous ESP process modeling methods, this study comprehensively considers the spatio-temporal coupling characteristics of ESP data and incorporates Crossformer into the dual-path contrastive learning (DCL) architecture to provide superior normal ESP process modeling. Furthermore, we design an ATG approach based on a random forest regressor that is aimed at successfully mitigating frequent false alarms resulting from fluctuations in ESP status. The algorithm is evaluated using data from four faulty wells in real oilfield scenarios, demonstrating its effectiveness and superiority through extensive comparative experiments against state-of-the-art methodologies.