A Deep Learning Approach for Heating and Cooling Equipment Monitoring

Condition monitoring is an important issue in system health management, and usually the first step leading to fault detection, diagnosis, and prognosis. It is often conducted through monitoring the time series of sensors, which usually includes outliers and change points. We adopt a deep learning model LSTM for monitoring the condition of boilers and chillers in a central heating and cooling plant. A two stage approach for condition monitoring is used: condition prediction and anomaly detection. In condition prediction stage, we use a LSTM model and three regression models: LASSO, SVR, and MLP to predict the energy efficiency of boilers and chillers. The experiments show that LSTM is able to establish a robust normal behavior of multiple boilers and chillers. LSTM reaches a lower prediction error than that of other three models in six out of nine boilers and chillers. In anomaly detection stage, we detect outliers or change points using the prediction errors from the LSTM model earlier than other models. This deep learning approach is applicable for real-time condition monitoring.