This paper deals with an fault detection and diagnosis (FDD) of appropriate refrigerant charge amount (RCA) using a feed-forward backpropagation neural network (FBNN) for multi-split variable refrigerant flow (VRF) systems. Faulty RCA operations of the VRF systems result in thermal discomfort for the occupants, lower coefficient of performance (COP), and equipment damage. Typical data driven neural network based methods give rise to computation complexity caused by data dimensionality and redundant data. Moreover, critical weakness of the BPNN results in deficient model generalization and over-fitting. This paper presents a fault detection scheme that uses the reliefF feature selection algorithm as a preprocessing technique to avoid the explosion of complexity while extraction critical feature information. Then, using a BPNN, it is shown that the proposed FDD algorithm renders the RCA of VRF systems classified. As a result the proposed technique can help to maintain the healthy VRF systems, provide thermal comfort, and save energy consumption.
[1]
M. Young.
The technical writer's handbook : writing with style and clarity
,
1989
.
[2]
Martin T. Hagan,et al.
Neural network design
,
1995
.
[3]
Jae-Bok Song,et al.
Torque sensor based robot arm control using disturbance observer
,
2010,
ICCAS 2010.
[4]
Marko Robnik-Sikonja,et al.
Theoretical and Empirical Analysis of ReliefF and RReliefF
,
2003,
Machine Learning.
[5]
F. Deyhimi,et al.
Prediction of activity coefficients at infinite dilution for organic solutes in ionic liquids by artificial neural network
,
2011
.
[6]
Beatriz de la Iglesia,et al.
Survey on Feature Selection
,
2015,
ArXiv.
[7]
Huanxin Chen,et al.
A refrigerant charge fault detection method for variable refrigerant flow (VRF) air-conditioning systems
,
2016
.