Anomaly Detection Based on Temporal Attention Network With Adaptive Threshold Adjustment for Electrical Submersible Pump

Accurate anomaly detection is critical for the electrical submersible pump (ESP) safety monitoring. Nevertheless, the multivariate, nonlinear, and dynamic nature of the ESP data poses significant challenges for this task. In this article, we propose a novel temporal attention network with adaptive threshold adjustment (TAN-ATA) to address the ESP anomaly detection problem. To model this multivariate nonlinear dynamic process, an encoder-decoder architecture based on long short-term memory (LSTM) is used as the backbone network in TAN. A temporal attention (TA) mechanism, in particular, is incorporated to reinforce the hidden state that contributes to boosting dynamic modeling performance. Furthermore, we propose an ATA strategy to combat the frequent false alarms caused by status fluctuations and employ a critical variable identification approach to locate root cause factors. Extensive experiments on practical data collected from a real Energy Development Company Ltd., China, illustrate the effectiveness and superiority of the proposed TAN-ATA method in terms of false alarm rate, lead time, and anomaly detection probability.