A Comparison between Polynomial and Locally Weighted Regression for Fault Detection and Diagnosis of HVAC Equipment

We investigate the accuracy of two predictive modeling methods for the purpose of fault detection and diagnosis (FDD) for HVAC equipment. The comparison is performed within an FDD framework consisting of two steps. In the first step, a predictive regression model is built to represent the dependence of the internal state variables of the HVAC device on the external driving influences, under normal operating conditions. This regression model obtained from training data is used to predict expected readings for state variables, and compute deviations from these readings under various abnormal conditions. The objective of the second step in the FDD framework is to learn to detect abnormalities based on regularities in computed deviations (residuals) from normal conditions. The accuracy of the first step (regression) is essential to the success of this method, since it disambiguates whether variations in observed state variables are due to faults or external driving conditions. In this paper, we present a comparison between locally weighted regression (a local non-linear model) and polynomial regression (a global non-linear model) in the context of fault detection and diagnosis of HVAC equipment. Our experimental results for detection and diagnosis of "overcharged" and "undercharged" refrigerant conditions in an HVAC device show that locally weighted regression clearly outperforms polynomial regression for this task