A successful fault detection and diagnosis (FDD) strategy for chillers is highly dependent on three core aspects, an accurate reference model, a sensitive fault detection technique and an effective fault identification method. An accurate model plays an important role in the performance of chiller FDD strategy. In this study, the accuracies of three models, MLR (multiple linear regression),,KRG (Kriging) and RBF (radial basis function), are compared under various modeling criteria. The EWMA (exponentially-weighted moving average) residual control chart, which is a more powerful tool in detecting small shifts than t-statistics based method, is adopted as a sensitive fault detection technique. The most sensitive feature parameter method is improved to effectively identify faults. Seven common faults in typical centrifugal chillers are considered. Three strategies (MLR-EWMA, KRG-EWMA and RBF-EWMA), which are the combinations of three reference models with EWMA and the most sensitive feature parameter method, are validated and compared by using two sample datasets with different sizes from ASHRAE RP-1043. Through the comparisons of diagnosis performance, the most accurate chiller reference model is found and suggested to detect and diagnose faults in chiller systems. The comparison results show that the RBF-based strategy has the best diagnosis performances among the three strategies. (C) 2016 Elsevier B.V. All rights reserved.