Application Of Hybrid-model Algorithm In Unknown Fault Diagnosis Of Air Conditioning System

Energy saving is one of the basic national policies for energy development in China.As the application scenarios of HVAC systems become wider and wider,their energy consumption expenditures are also becoming more serious.Due to its own or external reasons,HVAC systems often have faults and that causes unnecessary energy consumption during the operation process.Therefore,timely fault diagnosis is of great significance.The data-driven method can effectively use the system’s online and offline data to build models and realize diagnosis of unknown faults involved in the model training process,which has great potential in energy saving in HVAC.However,for those who did not participate in the model training process,it is unknown fault for the model,and the validity of the model can’t be guaranteed.Therefore,in this thesis,the diagnosis strategies are proposed for known faults and unknown faults.This thesis uses data of refrigerant charge fault obtained from experiments of VRF system to validate the proposed strategies.First of all,sufficient data are obtained from the experiments in the states of normal operation and fault operation to ensure data reliability for the model.Then,based on data change over time and outliers,some variables and samples are removed to reduce the invalid information contained in the data,and the input and output of the fault diagnosis model are determined.Finally,all data samples are divided into known faults and unknown faults for the model,then they are used to build training,validation,and testing dataset.Based on the inspiration of ensemble learning,this thesis proposes a hybrid-model algorithm.Since output of BP neural network is vector,it is employed for unknown fault diagnosis.At the same time,CART are employed for known fault diagnosis to reduce the misdiagnosis rate of known fault and improve the generalization performance.Firstly,BP neural network and CART models are built respectively by using the training dataset,and the diagnosis results of the two models are compared on the validation dataset.Then,based on BP neural network,a threshold-based strategy and a vector-similarity strategy are proposed to diagnose the unknown fault on the testing dataset.Finally,the hybrid model is used for integrated diagnosis,to improve the robustness of the model.The results show that on the validation dataset,the diagnostic accuracy of the CART decision tree is improved by nearly 10% compared to the BP neural network.The CART decision tree model possesses better diagnostic performance for known faults.On the testing dataset,the overall diagnostic accuracy of the threshold-based strategy and the vector-similarity strategy are 95.20% and 96.63% respectively,and the diagnostic accuracy of unknown fault are 94.47% and 97.44%.The two strategies realize the goal of unknown fault diagnosis without affecting the overall diagnostic performance of the model.