Refrigerant charge fault diagnosis in the VRF system using Bayesian artificial neural network combined with ReliefF filter

•Correlation between RCA and features is investigated through ReliefF algorithm.•Over 90% CDR can be achieved based on 22 features considering two indices.•The RCA fault can be diagnosed only based on 6-best features.•Training time of the model-6 is reduced by 98.8%. A proper refrigerant charge amou...

Full description

Saved in:
Bibliographic Details
Published in:Applied thermal engineering 2017-02, Vol.112, p.698-706
Main Authors: Shi, Shubiao, Li, Guannan, Chen, Huanxin, Liu, Jiangyan, Hu, Yunpeng, Xing, Lu, Hu, Wenju
Format: Article
Language:English
Subjects:
Artificial neural networks
Bayesian analysis
Bayesian neural network
Computational efficiency
Computing time
Fault diagnosis
HVAC
Neural networks
Refrigerant charge amount fault
Regularization
ReliefF algorithm
Residential buildings
Studies
Variable refrigerant flow system
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:•Correlation between RCA and features is investigated through ReliefF algorithm.•Over 90% CDR can be achieved based on 22 features considering two indices.•The RCA fault can be diagnosed only based on 6-best features.•Training time of the model-6 is reduced by 98.8%. A proper refrigerant charge amount (RCA) is critical for a variable refrigerant flow (VRF) system since RCA may affect the operational performance. However, there were few studies of RCA fault for the VRF system in the open literature. Therefore VRF systems are calling for a fault diagnosis strategy. This paper develops a highly efficient fault diagnosis model (FDM), which employs the ReliefF algorithm for feature ranking (FR) and applies the neural network for fault diagnosis. Firstly, the artificial neural network (ANN) model is built on the N-best features data subset and optimized by the Bayesian regularization algorithm. Secondly, the model is verified by testing data subset, the correct diagnosis rates (CDR) using the N-best features data subset can be obtained. The optimal FDM is selected in consideration of CDR and the computational efficiency. Finally, optimal FDM is further optimized by selecting the best hidden neurons. The results show that the CDR of the FDM based on 6-best features is sufficiently high in comparison to the CDR achieved when 22 features are used, while the training time decreases by 98.8%.
ISSN:1359-4311
1873-5606
DOI:10.1016/j.applthermaleng.2016.10.043