多联机制冷剂充注量故障检测模型评估及参数优化

Evaluation and Parameter Optimization of Fault Detection Model for Refrigerant Charge of Variable Refrigerant Flow System

本文对基于集成学习模型的多联机制冷剂充注量故障检测进行了评估,并对极致梯度提升(XGBoost)模型进行了参数优化。通过在制冷模式下进行制冷剂充注量故障实验采集了数据,结合互信息法和斯皮尔曼相关性系数法进行了特征选择,在此基础上分别训练了决策树、随机森林和XGBoost三种学习模型后得到诊断模型,并通过网格搜索等方法对XGBoost模型进行了参数寻优工作。结果表明:三种学习模型中基于随机森林的故障诊断模型效果最好,5个评估指标中精度由88.69%提高至93.46%、宏查准率由90.34%提高至98.09%、宏查全率由88.33%提高至98.19%、宏F1度量提高由88.94%提高至98.13%、kappa系数由0.87提高至0.98,而优化后XGBoost模型效果更优,各评价指标均再次提升1%~2%。

In this paper,the refrigerant charge fault detection of variable refrigerant flow(VRF)system based on the integrated learning model is evaluated,and the parameters of eXtreme Gradient Boosting(XGBoost)model are optimized.Through the refrigerant charge fault experiment in refrigeration mode,the data are collected,and the feature selection is carried out by combining mutual information method and Spearman correlation coefficient method.On this basis,the decision tree,random forest and XGBoost learning models are trained,and the parameters of XGBoost model are optimized by grid search method.The results show that,among the three learning models,the fault diagnosis model based on random forest is the best.Among the five evaluation indexes,the accuracy is improved from 88.69%to 93.46%,the macro precision is improved from 90.34%to 98.09%,the macro recall is improved from 88.33%to 98.19%,the macro F1 measurement is improved from 88.94%to 98.13%,and the kappa coefficient is improved from 0.87 to 0.98.After optimization,the effect of XGBoost model works better,and each evaluation index is increased by 1%-2%again.