The Study Of Fault Detection And Diagnosis For VaV System At Variable Conditions

The complexity of VAV air conditioning system leads to high frequencyof faults, and faults occurred to HAVC will cause energy losses, decline in airquality and comfort. With the widespread use of VAV air conditioning system,the study of its fault detection and diagnosis is getting more important.Currently, the researchers at home or abroad have obtained great achievementon study of HAVC. The fault diagnosis method can be divided into threecategories: model-based method, data-driven method and knowledge-basedmethod. But only one method can hardly achieve the detection and diagnosisof all the faults. The combination of these methods will achieve a better result.This study takes the air side of VAV air conditioning system as study object,and uses grey box model, clustering method, neural network for the faultdetection and diagnosis.Firstly, the process and state parameters of HVAC are affected by itsworking conditions. Hence, when using the model-based FDD method, thecomparison between prediction model and actual system has be conductedunder the same or similar working condition. Otherwise, the FDD result willbe invalid, even wrong. In this study, FCM clustering method is proposed toselect the fault-free reference data whose working condition is the nearest tothe current one, from large amounts of historical data, which is taken assample data for model training. Consequently, the accuracy of model can beimproved specifically, and the impact of working condition on FDD can beeliminated.Then, an AHU grey box model based on energy balance is built for faultdetection, which uses fault-free reference data for model training, and is used for getting predicted value. We can achieve the fault detection by analyzingresidual property between predicted value and actual measured value. TheAHU grey box model not only solves difficulty of modeling caused by highlynon-linearity and complexity of VAV air conditioning system, but alsoreflects physical properties of this system. Besides, the fault detectionthreshold is studied with statistical method, to apply to different workingconditions and faults.Thirdly, subtractive clustering method and fault classifier are combinedto find fault sources. The subtractive clustering method is used to judgewhether the current fault is trained fault in fault database. Then, thecharacteristics of common faults in Tsupcontrol loop are analyzed. And faultclassifier is built for determining fault type based on these faultcharacteristics.Finally, neural network prediction model is built to predict the residualof fault sensor for the purpose of fault tolerance, on the basis of relevanceanalysis between variables in Tsupcontrol loop.