Fault diagnosis of refrigerant charge based on PCA and decision tree for variable refrigerant flow systems

Conference Year

2016

Authors

Jiangyan Liu, Huazhong University of Science and Technology, Wuhan 430074, Hubei, ChinaFollow
Huanxin Chen, Huazhong University of Science and Technology, Wuhan 430074, Hubei, ChinaFollow
Jiangyu Wang, Huazhong University of Science and Technology, Wuhan 430074, Hubei, ChinaFollow
Guannan Li, Huazhong University of Science and Technology, Wuhan 430074, Hubei, ChinaFollow
Haorong Li, University of Nebraska-Lincoln, Lincoln, 68182, NE, USAFollow
Wenju Hu, Beijing Key Lab of Heating, Gas Supply, Ventilating andFollow

Keywords

fault diagnosis, refrigerant charge, decision tree, principal component analysis, variable refrigerant flow

Abstract

Variable refrigerant flow (VRF) systems are easily subjected to performance degradation due to refrigerant leakage, mechanical failure or improper maintenance after years of operation. Ideal VRF systems should equip with fault detection and diagnosis (FDD) program to sustain its normal operation. This paper presents the fault diagnosis method for refrigerant charge faults of variable refrigerant flow (VRF) systems. It is developed based on the principal component analysis (PCA) feature extraction method and the decision tree (DT) classification algorithm. Nine refrigerant charge schemes are implemented on the VRF system in the laboratory, which contain the normal and faulty refrigerant charge conditions. In addition, data of the online operating VRF systems are collected in this work. Firstly, data from both experimental VRF system and online operating systems are pre-processed by outlier cleaning, feature extraction and data normalization, because the original data of the VRF system usually has poor quality and complex structure. Secondly, the fault diagnosis model based on the PCA-DT method is built using the data of the experimental VRF system. In this step, the PCA method is used to obtain a new data sample which includes four comprehensive features, then the new data sample are randomly split into training and testing sets as the input of DT classifier for fault diagnosis. Thirdly, the advantages of the PCA-DT method is validated using the experimental data of different fault severity levels. Results show that the combined use of PCA and DT methods can achieve better fault diagnosis efficiency than the single decision tree method. Further, the robustness of the PCA-DT method in online fault diagnosis is verified using the data from online VRF systems. The online VRF systems have the same or different number of indoor units as the trained (experimental) VRF system. The PCA-DT method also shows desirable goodness on the online fault diagnosis process. In this sense, this work provides a promising fault diagnosis strategy for refrigerant charge faults of VRF system application.