Data-driven fault detection and diagnosis (FDD) for chillers depends heavily on the quantity of effective fault data (labelled). Plenty of labelled data is needed for a robust FDD model. Under data-scarce scenario when just a few labelled data are available, the model is easily biased. This study proposes a novel data pulling (DP) strategy based on a semi-supervised method to deal with this problem. Different from the traditional data cleaning idea, DP is concentrated on finding correct data to use instead of trying to identify all the noise to delete. System-level faults and component-level faults are treated separately and differently according to their distinct characteristics. Validations on seven types of chiller faults show that with scarce data, overfitting is easy to occur, random forest (RF) needs not be optimized, and the quality of the supplemented data is normally more important than the quantity. DP strategy can pull out data of high-confidence level, 99.87% for component-level faults and 100.00% for system-level faults. With just 20 labelled data for each fault, compared with the supervised model and the semi-supervised models without DP, the accuracy of the proposed method is improved by 1.77%∼4.63%, and the performance for refrigerant leakage is especially promoted.