Fault detection for centrifugal chillers using a Kernel Entropy Component Analysis (KECA) method

Fault detection is beneficial for chiller routine operation management in building automation systems.Considering the limitations of traditional principal component analysis(PCA)algorithm for chiller fault detection,a so-called kernel entropy component analysis(KECA)method has been developed and the development results are reported in this paper.Unlike traditional PCA,in KECA,the feature extraction or dimensionality reduction is implemented in a new space,called kernel feature space.The new space is nonlinearly related to the input space.The data set in the kernel feature space is projected onto a principal component subspace constructed by the feature space principal axes determined by the maximum Rényi entropy rather than the top eigenvalues.The proposed KECA is more suitable to deal with nonlinear process without Gaussian assumption.Using the available experimental data from ASHRAE RP-1043,seven typical chiller faults were tested by the proposed KECA method,and the results were compared to that of PCA.Two statistics,i.e.T2 and squared prediction error(SPE),were employed for fault detection monitoring.The fault detection results showed that the proposed KECA method had a better performance in terms of a higher detection accuracy in comparison to the traditional PCA.For the seven typical faults,the fault detection ratios were over 55%,even at their corresponding least severity level when using the proposed KECA based chiller fault detection method.

From laboratory to on-site operation:Reevaluation of empirically based electric water chiller models

Chiller model is a key factor to building energy simulation and chiller performance prediction.With spread of new types of electric water chillers that have higher performance and wider operating range,new challenges have been faced by building energy simulation tools and their chiller models.This work takes a new type of electric water chiller as a case study and reevaluates eight typical empirically based models for predicting the energy performance of electric water chiller to verify whether they are suitable for the new type of chiller,using both laboratory test data from chiller manufacturer and online monitoring data from on-site operation of a central cooling plant with chillers of the same type.The prediction ability of the chiller models(including model prediction accuracy and generation ability)in laboratory test and on-site operation situations are examined.The results show that the existing models can well describe the chiller performance in the laboratory test situation but perform poorly in the on-site operation situation.As the best two models in the laboratory dataset,the overall prediction errors of DOE-2 and GN model increase more than 250%and 75%respectively in the field dataset.The big discrepancy of model prediction accuracy in the two situations is mainly due to the differences of evaporator and condenser water flow rates between the laboratory and on-site operation datasets,which indicates the limitations of the empirical chiller models and implies further research in future in order to improve the suitability and reliability of chiller model.