Healthy sensors are essential for normal and optimal operation of HVAC&R systems.Various data-based methods are employed widely for sensor fault detection,diagnosis and error data reconstruction nowadays.Training data set from fielded historical operational data is carefully chosen for these data-based methods.This paper presents an optimal strategy based on cluster analysis to promote the quality of training data for PCA-based chiller detection,diagnosis and error data reconstruction.The presented training data optimal strategy can enhance the fault detection ratio and the data reconstruction accuracy after outliers are removed from the original data set.