Research On Fault Detection And Diagnosis Of Chillers Based On Bayesian Network

Chillers are the main energy consumption components in the heating,ventilation,air-conditioning(HVAC)systems in buildings.When chiller faults occur,they will shorten the life of equipment and cause energy waste.Fault detection and diagnosis(FDD)techniques can help people to find faults in the HVAC systems in time,which is of great significance in maintaining indoor environmental comfort,reducing equipment wear and tear and saving energy.Many FDD methods have been developed for chillers over the last decade,including PCA,ANN,SVM,SVDD,etc.Actually,the complexity of chiller faults,the limits to the data quantity and data acquisition accuracy make the mapping relations between fault features and faults become more random and uncertain.So it is reasonable to present FDD results in the form of probability,and probability theory can be used to solve this problem effectively.In this paper,we first introduce the chiller systems and analyze relationship between the typical faults and the symptoms.Then we build the chiller FDD model by using the Conditional Gaussian Network(CGN)and Discrete Bayesian Network(DBN),respectively,and give solution to the problems during their application process.For the CGN model,it realize the complete FDD process in a unique phase.However,the FAR is not in control when using the CGN model,and most users of HVAC systems can not accept the Higher FAR.So the Probabilistic Boundary is added in the CGN to keep the FAR in users’ acceptable range.At the same time,site information is added in the CGN to reduce the MDR of refrigerant overcharge(RO)fault by using the powerful ability of information fusion of BN.For the DBN model,in order to solve the problem that the time consumption of determining the parameters in the DBN by expert investigation is too much,we use the real distribution of the Normal operation data to determine thethreshold intervals of the state of the symptoms,and thees threshold intervals can help determine the parameters in CGN quickly.Site information is also added in the DBN to increase the CR of faults.Finally,we make an evaluation of this method on a 90-ton water-cooled centrifugal chiller reported in ASHRAE RP-1043.The results show that both the CGN and DBN achieve good results in chiller FDD process.For the CGN,the FAR decreases from 22.7% to 4.7% after integrating the probabilistic boundary,and the MDR of RO decreases from 27.6% and 9.4% after adding the site information.The CR of FDD are all over 90% except for the fault ConFoul.For the DBN,the CR are all over80% after adding the site information.