| The steady operation of the HVAC(Heating,Ventilation and Air Conditioning)system is of great significance to maintain the stability of the building’s thermal and humid environment.Various types of faults,especially those occurring in chillers,will not only cause the downgrade of HVAC system in controlling the indoor environment,increase energy consumption,but also shorten the lifespan of the equipment and cause high losses.The chiller is one of the most important equipment in HVAC system.At present,there are many research models dedicated to the fault detection and diagnosis(FDD)of chillers.However,there are also shortcomings such as subjective selection of modeling parameters,narrow use range and failure to predict faults in advance.On the basis of previous studies,this paper proposes a pre-diagnosis model for chiller’s FDD based on Support Vector Machine(SVM)and Grey Model(GM),which can predict whether a fault will come in the near future.The detection of the fault that may happen in the future is called a pre-diagnosis.The pre-diagnosis model of this paper consists of two parts,one is the FDD model based on SVM,and the other is the operational parameter prediction model based on GM.SVM has many advantages in tackling small sample and nonlinear classification problems,it can reach a high precision,calculation speed and in principle no overfitting problem.These characteristics form the basis of the FDD model.The result proves that we can use a joint diagnosis model to detect and diagnose multiple faults of the chiller;using the dimension reduction,noise reduction and filtering methods to process the sample data can improve the performance of the FDD model;only a small amount of core data can be used in the FDD of a larger amount of data;the fault diagnosis model established by the combination of Gaussian radial basis function and directed acyclic graph(RBF-DAG)has the most stable and best model performance;cross-system diagnosis of HVAC is feasible.The performance of GM using in onedimensional sequence prediction has been widely verified.It has the similar characteristic to SVM,and it only needs a small amount of data for modeling.When the coupling mechanism between parameters in the system is unknown,or the mathematical description is too complex,each parameter can be simply regarded as an independent gray parameter,and the development trend of the parameter can be directly predicted without considering the influence of other parameters.This idea provides a prerequisite for simplifying the prediction step.The results show that different singlefeature GM have different application scopes.For the same feature,when the prediction duration is set differently,the most applicable GM will not be same in most cases.Multi-feature GM considers the gray correlation between different parameters,so it is available to overcome the shortcomings of single-feature GM in prediction of some features and improve the performance of the pre-diagnosis model.Compared with other methods,the pre-diagnosis ideas proposed in this paper have more practical value in practical engineering than other studies that are more or less wise after the event. |
