Intelligent Diagnosis Strategies For Multiple Faults In Central Air Conditioning System And Their Transferring Application

The central air-conditioning system plays an important role in the building field.Once the system malfunctions,it will inevitable cause unnecessary energy consumption.Ensuring its efficient and stable operation is vital to building energy conservation.To this end,this dissertation takes the historical operating data of the system as the core,and utilizes data mining methods as the means to concentrate on the intelligent diagnosis strategies for the faults of central air-conditioning system and their transferring application.Taking the refrigerant charge fault diagnosis of the variable refrigerant flow(VRF)system as the starting point of research,it gradually evolves to the diagnosis of multiple types of faults in different systems.Continuous exploration and innovations have been carried out from both feature selection methods and fault diagnosis strategies.The application of the ensemble learning idea solves the problem of poor stability of the diagnostic model and simultaneous diagnosis of multiple types of faults,and the introducing of transfer learning method realizes the migration of fault diagnosis models between different systems under the condition of lack of fault data.Finally,a complete intelligent fault diagnosis system for building central air-conditioning system has been constructed.The main research contents of this dissertation are as follows:Aiming at the diagnosis of the most common refrigerant charge faults in VRF central air-conditioning systems,a weighted K-nearest neighbor fault diagnosis model based on coupling feature selection is proposed.By coupling multiple single feature selection methods,a better set of critical fault characterization variables is obtained.The weighted Knearest neighbor model established based on the feature set shows high diagnostic accuracy for refrigerant leakage and overcharge faults corresponding to different severity levels,and its effectiveness has been well verified on the other four different types of VRF systems.In order to more accurately diagnose the amount of refrigerant remaining in the system,an automatic optimization diagnosis strategy for refrigerant charge amount faults is further proposed,and the strategy can effectively avoid the abnormal diagnosis performance caused by the difference in the principle of the diagnosis model,and diagnose the refrigerant remaining in the system with higher accuracy.Aiming at the problem of poor stability and versatility of a single fault diagnosis model,a two-stage ensemble diagnosis model is proposed.The simulated annealing algorithm is used to select the key characterization variables of refrigerant charge amount faults,valves faults and compressor faults.In the first stage of the integration process,the boosting method is adopted to integrate multiple component learners,which effectively improves the diagnostic performance of the preliminary ensemble models.In order to further improve the stability and versatility of the model,a weighted voting strategy is finally adopted to integrate the results of the preliminary ensemble model.The results show that this process can effectively integrate the advantages of each preliminary integrated model.This twostage ensemble model has good diagnostic performance and stability in the diagnosis process of refrigerant charge faults,valve faults and compressor faults in VRF system,perfectly avoiding the defect that the performance of a single fault diagnosis model may be greatly degraded due to different application objects.Aiming at the problems that it is difficult to diagnose multiple types of faults efficiently simultaneously and different feature selection methods tend to obtain the local optimal key characterization variables of the fault,a Light GBM multi-faults diagnosis model combined with the sequence forward search ensemble feature selection method is proposed.Innovatively introducing the ensemble learning idea into the selection process of the key fault characterization variables for multiple faults.First,five basic feature selection methods are used to obtain feature subsets arranged in descending order of importance,and then the sequence forward search strategy is used to search for the optimal feature subset in the feature space composed of the five feature subsets.And the diagnostic performance of the Light GBM model is served as a criterion for evaluating the quality of features,which can effectively avoid obtaining the local optimal feature subset,and the finally selected feature subset can better characterize the symptoms of various faults.The results show that the Light GBM diagnostic model based on the ensemble feature selection method can diagnose eight common faults of the VRF central air-conditioning system at the same time.The effectiveness and versatility of this method have been well verified in the diagnosis of seven types of typical fault in chillers.Finally,in view of the difficulty of establishing a central air-conditioning system fault diagnosis model under the condition of lack of fault data,this dissertation proposes a fault diagnosis strategy based on the transfer learning method.First,four common faults of the central air-conditioning system are simulated and the operating data of the simulation system under different application scenarios are obtained.Based on this,the multi-faults diagnosis performance of two transfer learning methods in different application scenarios is studied.By using the pre-training model as the transferring method of the feature extractor,the bottleneck features of the central air-conditioning system of the target building with insufficient fault data can be effectively extracted.The transfer learning method that finetunes the pre-training model can better learn the information of the target building while retaining the knowledge learned from the source architecture.The central air-conditioning system fault diagnosis model based on the two transferring methods possess excellent diagnostic performance in different scenarios,effectively solving the problem that it is difficult to establish an efficient diagnosis model due to the lack of fault data,and realizing the faults diagnosis model transferring of different air-conditioning systems in different buildings,having great application potential.This dissertation proposes a coupled feature selection method to improve the diagnosis accuracy and efficiency of diagnosis model for refrigerant charge amount faults in the VRF system.On this basis,the factors affecting the stability of the fault diagnosis model are analyzed,and two-stage ensemble diagnosis model suitable for different types of faults is proposed.Furthermore,the strategy of simultaneous diagnosis of multiple faults is further studied,and the ensemble learning idea is successfully applied to the selection process of key characterization variables of multiple faults.Aiming at the problem of multi-fault diagnosis model transferring between different air-conditioning systems in different buildings,a multi-fault diagnosis model based on the transfer learning method is established,which improves the multi-fault intelligent diagnosis frame of central air-conditioning system in building.