Health Estimation and Fault Prediction of the Sensors of a HVAC System

Abstract

Due to ageing or adverse environment, the sensors of an HVAC system deteriorate progressively and fail to produce the desired output after sometime. Each HVAC system has hundreds of sensors. This paper proposes a generic framework to predict the failures of these sensors in advance. A novel technique has been used to transform the problem domain from prediction to detection where conventional algorithms were used to build classifiers. A number of common features were derived out of the sensor values. These features were subsequently used to define a function to deduce in real time the health of a sensor. A dashboard displays the deterioration of the health of the sensor over the time. Data from hundreds of sensors of more than 60 HVAC systems with hundreds of sensors each were used to build machine learning models. The solution has been deployed to detect failure of these sensors and it was found that this framework was able to model 74% of all sensor faults at least 10 hours in advance. The accuracy of fault prediction has been more than 96%, precision has been more than 74% and recall has been 95%.