【摘要】 冷水机组故障导致中央空调系统运行效率降低，产生较大能耗浪费。冷水机组故障诊断普遍存在诊断精度较低、诊断速度较慢、诊断模型更新用时较长且可解释性较差的问题。本文提出了一种基于两步特征选择和贝叶斯优化LightGBM的复合故障诊断新方法。首先，采用嵌入法和递归特征消除法相结合的两步特征选择方法，对关键故障特征进行提取；其次，将LightGBM算法运用于冷水机组故障诊断中，并利用贝叶斯优化算法对故障诊断模型参数进行离线优化；最后，采用ASHARE RP-1043项目数据验证本文提出的方法。实验结果表明：本文所提特征选择方法可有效识别出冷水机组的关键故障特征，所选5个特征均属于制冷系统的重要参数，在所选特征维度下，构建的LightGBM模型比嵌入法更优，模型性能提升约5.12%;相比于贝叶斯优化后的XGBoost和支持向量机(SVM)模型，贝叶斯优化后的LightGBM模型总体精度和Kappa系数更优，训练样本量较小时诊断精度相对更高，且其训练速度更快、诊断用时更短，训练样本量较大时训练速度和诊断速度优势更为明显，当训练样本量从5 000增加至50 000,模型诊断总体精度从96.11%提高到99.40%。此外，在LightGBM模型上，将贝叶斯优化与粒子群优化和遗传算法进行性能对比，验证了贝叶斯优化在参数寻优上具有优势。本文研究结果可为大数据背景下冷水机组的故障诊断提供理论指导和依据，有助于管控冷源系统的运行风险。更多还原

【Abstract】 The chiller faults reduces the operation efficiency of the central air conditioning system and produces large energy waste. The fault diagnosis of chiller usually comes with low diagnosis accuracy, slow diagnosis speed, long diagnosis model update time and poor interpretability. To this end, a new hybrid fault diagnosis method based on two-step feature selection and LightGBM with Bayesian Optimization was proposed. A two-step feature selection method combining embedded and recursive feature elimination was first used to extract key fault features. Then, the LightGBM algorithm was applied to chiller fault diagnosis, and the fault diagnosis model parameters were optimized offline based on Bayesian optimization. This method was further verified with the data provided by the ASHARE RP-1043 project. The experimental results show that the proposed feature selection method can effectively identify key fault features of chillers, and the five selected features are important parameters of the refrigeration system. The model performance is improved by about 5.12% compared with LightGBM constructed by embedded under selected feature dimensions. The overall accuracy and Kappa coefficient of LightGBM with Bayesian optimization, are superior to those of XGBoost with Bayesian optimization and support vector machine(SVM) with Bayesian optimization. The proposed model has faster training speed and shorter diagnostic time, especially when the sample size is smaller, its diagnosis accuracy is relatively higher, and when the sample size is larger, the advantage of its training speed and diagnostic speed is more obvious. Moreover, when the training sample size is increased from 5,000 to 50,000, the overall accuracy of the proposed model is improved from 96.11% to 99.40%. Further, the performance among Bayesian optimization, particle swarm optimization and genetic algorithm was compared on LightGBM to verify the advantage of Bayesian optimization in parameter optimization. The research results can provide theoretical guidance and basis for chiller fault diagnosis in the context of big data, and help to control the operational risk of the cold source system.更多还原