Application Of Ensemble Learning In The Refrigerant Charge Fault Diagnosis For Variable Refrigerant Flow Air Conditioning System

In recent years,Variable refrigerant flow(VRF)air conditioning system has been widely used in all kinds of public buildings.The refrigerant charge level is an important parameter that affects the efficient operation of air-conditioning system.In this paper,a fault detection and diagnosis strategy based on ensemble leaning is proposed by taking the fault of VRF refrigerant charge as an example.The VRF air conditioning system operates at different refrigerant charge levels.The original data set collected by the experimental device contains 67336 samples and 216 characteristic variables under normal and fault conditions.According to the refrigerant charge range and sample distribution,the refrigerant charge data is divided into 5 categories,and corresponding to the labels L-2,L-1,L0,L1 and L2.In order to improve the quality of modeling data,pre-process the refrigerant charge data,make preliminary feature selection of feature variables,use box-plot method to remove abnormal data samples,and standardize to eliminate dimensional differences.The fault diagnosis strategy proposed in this paper uses the principal component analysis method and chi-square test algorithm for further feature processing of the preprocessed refrigerant charge data,combined with the fault diagnosis results of random forest to select the feature subset with the best prediction performance.Finally,build an integrated fault diagnosis model based on Boosting,and make a comparative analysis with the diagnosis results of the three basic classifiers of C4.5 decision tree,support vector machine,and BP neural network.The overall correct rate(CR)and individual hit rate(HR)are used as model evaluation indexes.The results show that the principal component analysis method is more suitable for the feature processing of VRF refrigerant charge fault detection and diagnosis than the chisquare test algorithm.Taking into account the cumulative contribution of variation of the principal components,the training time and the prediction accuracy of the random forest,the optimal input feature subset selects 6 new principal variable data sets obtained by the principal component analysis method.Compared with the diagnostic results of the three basic classifiers and the random forest classifier,the Boosting ensemble model has the best effect,and the predicted CR is 94.86%,and the HR for each refrigerant charge failure category is Above 90%.