Fault detection and diagnosis of chiller using Bayesian network classifier with probabilistic boundary

Abstract

False alarm rate (FAR) is an important evaluating indicator when chiller fault detection and diagnosis (FDD) system is applied on site, and higher FAR will not be accepted by users and manufacturers. When we use Bayesian network classifier (BNC) to operate FDD process, it is hard to keep the FAR within an acceptable range when the class node contained the normal operating condition. So a practical chiller FDD methodology is proposed. At first, a probabilistic boundary limit under the normal operating condition is integrated in the BNC to reduce the FAR in chiller FDD process. Then, we add the site information to the BNC to solve the problem that missed detection rate (MDR) rises caused by the reduction of FAR. Evaluation of the methodology was made on a 90-ton water-cooled centrifugal chiller reported in ASHRAE RP-1043. The results show that the FAR decreases from 22.7% to 4.7% after integrating the probabilistic boundary, and the MDR of refrigerant overcharge and refrigerant leakage decreases from 27.6% and 9.4–7.6% and 4.5% respectively after adding the site information.