Data Driven Approach-based Fault Detection And Diagnosis For The VRF Air-conditioning System

The fault of the heating,ventilation and air conditioning(HVAC)system may cause the equipment performance decline or failure,which will lead to a series of problems such as energy waste,reduced indoor thermal comfort and reduced equipment service life.Therefore,this thesis focuses on the fault detection and diagnosis of the air conditioning system.Due to characteristics and advantages of the variable refrigerant flow(VRF)system,it has been widely used in small and medium-sized buildings in recent years,but there are few researches on fault detection and diagnosis of the VRF system.Therefore,the research object of this study is focus on the VRF system.Because the VRF system is highly automated,and the equipment itself has been installed a large number of sensors.In addition,some VRF systems have data colletion and transmission capabilities.Therefore,it lays a data foundation for the fault detection and diagnosis study of the VRF air-conditioning system using data-driven methods.Moreover,machine learning and data mining methods have been rapidly developed in image recognition and speech recognition fields.Therefore,the data-driven method is an effective way to slove the fault detection and diagnosis of air-conditioning systems.This thesis focuses on sensor faults,indoor unit faults and outdoor unit faults of the VRF system.Aiming at the problems of operational data fluctuation,large number of indoor units,feature variable selection and deep learning application,a series of fault detection and diagnosis studies are carried out by using data-driven approaches,including SG method,principal component analysis(PCA),association rule mining,neural netword and deep belief network(DBN).The main research contents are as follows:Since the data collected by the VRF system has zigzag fluctuations,it will affect the sensor fault detection and diagnosis performance.This thesis proposes a data smoothing optimization method based on SG algorithm to optimize the data,then uses the PCA method to establish the SG-PCA sensor fault detection and diagnosis model.In order to determine the two parameter settings of polynomial order and sliding window width in SG method,a parameter optimization index is proposed,which is calculated by signal-to-noise ratio,standard deviation and self-detection efficiency of the smoothing data.Then the faults with different severity are introduced into the actual running data to verify the effect of the PCA model.The results show that compared with the traditional PCA model,the SG-PCA model can significantly improve the sensor fault detection and diagnosis performance.Aiming at the characteristics of the VRF system with multiple indoor units,this thesis proposes a modularized PCA model fault detection strategy.This strategy can not only detect the fault,but also determine which indoor unit is fault.Firstly,the modularized PCA method is used to establish fault detection models for indoor and outdoor units in various modes.Then multivariable decoupling fault diagnosis strategy based on expert knowledge is developed by using six feature variables.Four faults have been used to evaluate the fault detection and diagnosis strategy proposed in this study,including two indoor unit faults(electronic expansion valve fault and indoor heat exchanger fouling),one outdoor unit fault(four-way reversing valve fault)and a system level fault(refrigerant undercharge).The result analysis shows that the modularized PCA fault detection method and multivariable decoupling fault diagnosis strategy are effective for the above four faults.In order to slove the problem of feature variable optimization selection,a feature selection method based on data mining method for fault diagnosis model is proposed.Firstly,the correlation analysis method is used to remove the redundant variables,and then the association rule mining method is used to optimize feature variable selection.Through feature optimization,a total of five feature variable sets(FS1-FS5)are selected.The FS1 is the original feature set.The FS2 is feature set after removing redundant variables,and FS3-FS5 are the feature set mined by association rules.The fault diagnosis models based on neural netword method are established by using above feature sets,and the performance of models is evaluated by four kinds of fautls: outdoor unit heat exchanger fouling,four-way reversing valve fault,refrigerant undercharge and overcharge faults.The results show that the correlation analysis method can effectively eliminate redundant variables,and the association rule mining method can be used to optimize the feature variable set.The fault diagnosis model with FS5 has the best fault diagnosis performance,and the fault diagnosis correct rate has increased from 88.71% to 96.40%,and the hit rates of all four faults are higher than 90%.Finally,this thesis proposes a fault diagnosis strategy based on deep learning method(deep belief network),and its application potential in the field of air conditioning system fault diagnosis is studied.For the deep learning method with a lot of parameters and difficult selection problems,this thesis analyzes the influences of each parameter on the performance of the DBN model,and then proposes a parameter optimization selection strategy.The DBN mdoel is evaluated by four faults in heating mode.The results show that the deep belief network model can be applied to the fault diagnosis of air conditioning systems.Through parameters optimization selection,the fault diagnosis correct rate of the optimized model reaches 97.70%,an increase of 5.05% over the DBN model with initial parameters.In addition,the depth of DBN mdoel is optimally selected as 2 layers,indicating that the DBN mdoel may not need to be very deep for the VRF system fault diagnosis.Besides,the performance of the DBN model and the above neural network model are compared and analyzed.The results show that the DBN mdoel has better fault diagnosis performance.