Fault Detection for Circulating Water Pump Using Time Series Forecasting and Outlier Detection

In any power plants, it is crucial to perform a preventive maintenance to avoid unexpected breakdown of machinery, e.g., circulating water pump, using data collected from various sensors. There have been prior attempts using just traditional prediction techniques. In this paper, we propose a two-stage model that employs a technique from time series analysis to predict when the machine tends to be failed for one day in advance. The first stage focuses on forecasting trends of each sensor using "Auto-Regression Integrated Moving Average (ARIMA)." Then, the second stage aims to classify failure mode using the predicted sensor values. The experiment was conducted on data collected from eight sensors within one year. The result is shown that our proposed algorithm significantly outperforms an existing technique, Regression Artificial Neural Network.