There are differences among different levels of the same type of the fault

which may cause misdiagnose. A fault diagnosis strategy based on multi-scale principal component analysis and kernel entropy component analysis (MSPCA-KECA) is proposed. Taking the features extracted by MSPCA as the input of KECA classifier can be used for fault online detection as well as automatic identification. MSPCA combines wavelet multi-scale analysis with principal component analysis to select the scales which contain fault-related information

and then use PCA to extract the fault-related features

extracting the similarity among different levels of the same type of fault and the difference among different faults

which can improve the ability of fault diagnosis. The combination of KECA and Cauchy-Schwarz (CS) statistics extract and express the angular structure of different kinds of faults

which is good for fault classification. The control limit here is achieved by support vector data description (SVDD) for the unacquainted distribution of the statistics. Through the simulation of ASHRAR 1043-RP chiller data

the feasibility and effectiveness of the MSPCA-KECA method are verified.