Feature Selection for Chiller Fault Detection and Diagnosis Based on Grey Similitude Degree of Incidence.

From the perspective of online applications of chiller fault detection and diagnosis (FDD) systems, a feature selection (FS) method based on grey similitude degree of incidence (GSDI) is proposed, in order not only to save initial sensor costs of online applications through reducing the number of sensors and selecting low-cost sensors but also to keep the excellent FDD performance. Firstly, under the constraints of online applications: low cost of measurement and high sensitivity to faults, 16 candidate features are determined. Secondly, the optimal number of features is determined based on mutual information (MI) technique. Thirdly, the specific feature subsets are determined based on GSDI. Finally, a multi-class support vector machine (SVM) is introduced as an evaluation tool to evaluate the FS results through its diagnostic performance. Seven typical faults of the chiller are concerned in this paper. The experimental data are from the ASHRAE RP-1043. The results show that the number of features is reduced effectively, meanwhile keeping the excellent FDD performance, i.e., the diagnostic correct rates exceed 95%; the misdiagnosis rates are lower than 0.5%; the false alarm rates are lower than 1%. [ABSTRACT FROM AUTHOR]