Research On Fault Diagnosis Method Of HVAC System Based On Feature Selection

Heating Ventilation and Air Conditioning(HVAC)systems are widely used in various buildings,providing the necessary guarantee for comfortable life of people.The HVAC system has a complex structure and consumes a lot of energy.The annual energy consumption of HVAC systems accounts for more than 40% of the global building energy consumption,and the proportion is increasing rapidly.HVAC systems consume more energy while operating under fault conditions.Studies have shown that maintaining the normal operation of HVAC systems can save 30% of the total building energy consumption.Therefore,Researchers developed a large number of fault detection and diagnosis methods for HVAC systems to maintain the energy efficiency of buildings.Thanks to the large amount of historical data collected in building automation systems and the development of machine learning,data-driven FDD methods have gradually gained favor among researchers.However,in general,HVAC systems run in a fault state for less than the normal state,which also results in less fault data,and the lack of faulty samples in the training process remains as a difficulty for FFD of HVAC systems.This thesis proposes a method for fault diagnosis of HVAC systems based on feature selection.This method first obtains a set of low redundancy features,and then uses a small amount of data to train a learner based on these features.Finally,use the trained learner to complete the FDD of the HVAC system.This method can obtain high accuracy of fault diagnosis with only a small amount of training data.Specifically,this thesis first uses a spectral clustering algorithm to cluster the time series scattered in each sub-feature space,and then summarizes the results in the sub-feature space to obtain the total clustering result.Based on the entropy value of different time series,the initial feature set is obtained from the clustering results and the clustering clusters are sorted,and then the feature selection is performed to obtain the result set.After that,this thesis train the learner based on the feature set and used for FDD.In order to verify the performance of the FDD algorithm proposed in this thesis,this thesis uses the HVAC system operation data set collected from the real world to test the algorithm proposed in this thesis.The experimental results show that the method in this thesis has better generalization ability in fault diagnosis.When the training data accounts for 2% of the total data,it can obtain a diagnosis accuracy rate of more than 90%.Compared with the control group,it can achieve a maximum improvement of more than 30%.When the training data accounts for 5% of the total data,the method in this thesis can obtain a diagnosis accuracy of 99.3%.The above experiments show that the method proposed in this thesis can solve the problem of too few fault data better.