In commercial buildings, heating, ventilation, and air-conditioning systems account for nearly 40% of total energy consumption, with chillers serving as the core component; thus, accurate chiller fault diagnosis is essential for reducing energy waste. However, conventional diagnostic methods typically assume balanced datasets, while in practice fault samples are far fewer than normal samples. To address this issue, this study develops an integrated diagnostic framework. First, key features are selected by combining RF, CatBoost, and Pearson correlation analysis. Subsequently, ADASYN is employed to alleviate class imbalance, and a one dimensional convolutional neural network is employed to extract representative features from the augmented data. On this basis, a cost-sensitive XGBoost diagnostic model that integrates class weights and cluster weights is developed to enhance the model’s attention to minority fault classes. Finally, SHAP is employed to interpret the model’s decision-making process, thereby improving model interpretability. Extensive experiments are conducted on the ASHRAE RP-1043 dataset at the mildest fault severity level under imbalance ratios of 20:1, 50:1, 100:1, 200:1, and 300:1. For example, under the 300:1 imbalance ratio, each fault class contains only 12 samples, while the proposed framework achieves average accuracies of 98.18%, 94.96%, 90.94%, 85.57%, and 80.39%, respectively. Comparative results demonstrate that the proposed method achieves superior fault diagnosis performance compared with existing methods.