Anomaly detection of a cooling water pump of a power plant based on its virtual digital twin constructed with deep learning techniques

Abstract

This paper aims to explore the use of recent approaches of deep learning techniques for anomaly detection of potential failure modes in a cooling water pump working in a gas-combined cycle in a power plant. Two different deep learning techniques have been tested: neural networks and reinforcement learning. Two virtual digital twins were developed with each family of deep learning techniques, able to simulate the behavior of the cooling water pump in the absence of pump failure modes. Each virtual digital twin consists of several models for predicting the expected evolution of significant behavior variables when no anomalies exist. Examples of these variables are bearing temperatures or vibrations in different pump locations. All the data used comes from the SCADA system. The main features and hyperparameters in the virtual digital twins are presented, and demonstration examples are included.This paper aims to explore the use of recent approaches of deep learning techniques for anomaly detection of potential failure modes in a cooling water pump working in a gas-combined cycle in a power plant. Two different deep learning techniques have been tested: neural networks and reinforcement learning. Two virtual digital twins were developed with each family of deep learning techniques, able to simulate the behavior of the cooling water pump in the absence of pump failure modes. Each virtual digital twin consists of several models for predicting the expected evolution of significant behavior variables when no anomalies exist. Examples of these variables are bearing temperatures or vibrations in different pump locations. All the data used comes from the SCADA system. The main features and hyperparameters in the virtual digital twins are presented, and demonstration examples are included.