Research On Fault Pre-diagnosis Method Of Chiller Based On Bayesian Network

Faults of chillers will affect indoor thermal comfort,and increase system energy consumption.Pre-diagnosis can predict the faults that going to happen timely and effective,which is beneficial to the inspection and elimination of potential hazards.The research of chiller fault pre-diagnosis is an important way of reducing operation and maintenance costs,keeping the chillers high-efficiency operation and saving energy of the whole HVAC system.In the past few decades,a lot of research work on fault detection and diagnosis have been done,but there is little research on pre-diagnosis.With the further improvement of the system security and reliability requirements,people not only hope to diagnose system faults after they occur,but also hope to predict the fault when only minor abnormal signs appear,that is,to judge the trend of fault development and the degree of equipment performance degradation,which can avoid major economic losses and safety accidents.In the study of fault pre-diagnosis,the method of expert system,fuzzy logic,neural network,etc.,have been proposed.However,in view of the complexity of chiller system and fault,the high dimension and non-linearity of sensor data,the uncertainty of the mapping relationship between fault and symptom,etc.,Bayesian network(BN)is adopted to realize fault pre-diagnosis of chiller in this study.BN is a powerful tool for the expression of uncertain knowledge,with good flexibility and the ability to integrate all kinds of valuable information.Therefore,BN has great advantages and good application prospects in the field of fault pre-diagnosis.BN is not easy to deal with continuous values,and the gaussian hypothesis is not always consistent with the reality.Therefore,the discretization technology which is used for discrete continuous values is introduced to the fault pre-diagnosis based on BN in this study.Firstly,this paper introduces the chiller system and typical faults.Secondly,four typical discretization algorithms are compared and analyzed.Then,the fault development trend prediction model and the equipment performance degradation prediction model that based on discrete BN are established.At the same time,for the model with low accuracy,it can be improved to a higher level by combining manual observation information.Finally,ASHRAE RP-1043 data is used to verify the pre-diagnosis model of the chiller.The results show that,in the constructed fault development trend prediction models,the discrete BN prediction model based on the minimum description length(MDLP-BN)has the best effect,and its average prediction accuracy is 75%,higher than other methods,which can realize effective prediction.The proposed prediction model of equipment performance degradation degree has a good performance,for which the prediction accuracies are more than 85% under the three fault modes of Reduced condenser water flow(Redu CF),Reduced evaporator water flow(Redu EF)and non-condensable gas(Non Con).