应用于冷水机组的故障诊断技术对于降低建筑能耗，提高机组运行效率有着重要作用。为了进一步提高冷水机组故障诊断性能，同时考虑到特征参数残差蕴含更多故障信息，提出一种基于特征参数残差驱动贝叶斯网络(BN)的冷水机组故障诊断方法。首先，构建特征参数基准值模型，获得其基准值；然后使用基准值和其实际运行值之间的参数残差训练BN模型。以此达到充分利用参数残差蕴含故障信息，从而提高故障诊断性能的目的。使用实验数据对构建方法的有效性进行验证，与使用特征参数直接驱动BN的诊断模型相比，对冷水机组常见故障的诊断正确率最高提升了约22.51个百分点。此外，比较分析了三种参数基准值模型构建方法的性能，基于神经网络方法的基准值模型较其他两种基准值模型表现更优。

Chillers, as a major consumer, account for about 40% of the total building energy consumption. However, the failure of chillers will lead to additional energy waste accounting for 15%~30% of building energy consumption. Therefore, applying fault diagnosis technology to chiller plays an important role in reducing energy consumption and improving operation efficiency. On the other hand, considering that the residuals of parameters involve more information reflecting faults, a fault diagnosis method based on parameter residual-driven Bayesian Network (BN) is proposed by combining parameter residuals with BN for chiller, in order to furtherly improve the fault diagnosis performance. Being different from most of the conventional methods directly using the parameter measurement values to train the models, the proposed method uses the residuals, calculated through the actual values and reference values of parameters to train the BN model, thus to make full use of the fault information contained in parameter residuals. To evaluate comprehensively the effectiveness of parameter residuals referring to enhance the diagnostic performance, three models used to determine the reference values are developed and compared. Two are linear analysis methods, i.e., multivariate linear regression (MLR) and partial least squares regression (PLSR), and the other is a nonlinear analysis method, i.e., back-propagation neural network (BPNN). Finally, the proposed method based on parameter residual-driven BN is applied to a real experimental chiller, and the experimental data are used to verify its effectiveness. The results show that: (1) Compared with the diagnosis model driven by parameter measurement values directly, the proposed method has higher diagnosis accuracies for the considered seven common faults of chiller, and the diagnosis accuracy is increased by 22.51 percentage point at most; (2) Compared with the reference model based on MLR and PLSR, the diagnosis performance is better when the BPNN model is used to determine the reference value, and the diagnosis accuracy is increased by 12.35 and 12.05 percentage point at most, respectively; (3) The proposed method can effectively improve the diagnostic performance, especially for these faults at slight severity level.