Incipient fault detection and diagnosis for centrifugal chillers is significant for maintaining safe and effective system operation. Due to the advantages of simple learning algorithm and high generalization capability, the extreme learning machine (ELM) can identify faults quickly and precisely in comparison to conventional classification methods such as back propagation neural network (BPNN). This paper reports an effective diagnosis method for incipient chiller faults with the integration of kernel entropy component analysis (KECA) and voting based ELM (VELM). KECA was first performed to reduce the dimensionality of the original input data so as to minimize the model complexity and computational cost. Instead of using a single ELM, multiple independent ELMs were adopted in VELM, and then the class label could be predicted based on the majority voting method. Using the experimental data of seven typical faults together with a normal operation, the proposed KECA-VELM fault diagnostic model was trained and further validated. The results show that a better fault diagnosis performance can be achieved using the KECA-VELM classifier compared with the conventional BPNN, ELM and VELM based classifiers. The overall average fault diagnosis accuracy for the faults at the least severity level was reported over 95% based on the proposed method.