Experimental And Theoretical Research On Integrated Fault Detection And Diagnosis Model For Two Faults In VRF System Based On 1-D CNN

According to incomplete statistics,China’s air-conditioning energy consumption has accounted for 50～65%of the total building energy consumption.VRF(variable refrigerant flow)system is one of the main forms of centralized air conditioning systems.Establishing an accurate and efficient FDD(fault detect and diagnosis)model for VRF air-conditioning systems is of great significance for reducing the energy consumption of VRF systems.Therefore,this paper explores the establishment of an integrated model for multiple FDD of VRF systems.Based on the deep learning method 1-D CNN(one-dimensional convolution neural network),an integrated model for FDD of two typical VRF system faults is established,and the model is analyzed and evaluated.The integrated model in this study is proposed based on experiment of gas-liquid separator opposite-insertion and refrigerant charge fault of VRF system under heating conditions.The experiment was conducted in the enthalpy difference laboratory to collect the fault data and normal data of the VRF system under the two system fault conditions,so as to study the double fault detection and diagnostic modeling from the experimental and theoretical perspectives.Data preprocessing is conducted in this study,which includes data cleaning,normalization,labeling design,and feature selection.The cleaned data set includes three types of fault states that need to be detected:refrigerant charge fault,normal and gas-liquid separator opposite-insertion fault.It also includes six levels of refrigerant charge fault,namely L-3 to L+3,that need to be diagnosed.The integrated model establishes two sub-modules,namely1-D CNN_aand 1-D CNN_b,whose parameters and network structures are specifically adjusted to suit the VRF system operation data.The train set and test set are divided into 75%and 25%for training and testing.1-D CNN_aand 1-D CNN_bare trained separately on the train set and jointly tested on the test set.The evaluation results using multiple evaluation indicators and comparison models show that the integrated model based on 1-D CNN achieves better fault detection,double fault detection and double fault diagnosis results than the comparison models that based on algorithms such as BPNN,DT,DBN,and SVM.Grid search method was then used for the 1-D CNN_aand 1-D CNN_bof the integrated model to optimize the initial learning rate,the number of batch samples and the number of training epochs,which further improved the performance of the integreted model.Based on the analysis of fault detection and diagnosis results,this study summarizes the experience of the double fault detection and diagnosis problem and modeling methods.According to the results obtained in this study,the FDD of refrigerant charge charge fault detection and diagnosis is more difficult than the gas-liquid separator opposite-insertion.The results also show that for VRF system fault detection and diagnosis model problems,dividing the problem reasonably,choosing algorithms wisely,and selecting features with universal advantages while highlighting features that have special effects on individual problems are the keys to improve the overall performance of the integrated FDD model.This study established an integrated 1-D CNN FDD model for two typical fault in VRF system,and gained some experience in modeling methods,which lays a foundation for the future research to establish integrated models for more general FDD in VRF systems.