Research On The T-SNE And GRU Network Based Fault Diagnosis Method For Vapor Compression Refrigeration Systems

It was reported in open literature,China’s building energy consumption accounts for about 30%of all social energy end-use,while HVAC accounts for more than 40% of building energy consumption,More than 20% of energy will be consumed,if air-conditioning equipment "Operation with failure".If these faults can be detected and eliminated in time,energy consumption can be greatly reduced.Therefore,an increasingly number of research have been carried out to the fault detection and diagnosis of HVAC systems.This project aims to find a fault detection and diagnosis method for refrigeration systems based on data-driven.A neural network fault diagnosis method based on distance characteristics is proposed,verified by the experimental data of ASHRAR RP-1043 and a real experimental heat pump simulation and experiment.The specific work is as follows:(1)Generally,Euclidean distance is widely applied in fault detection feature extraction.However,Euclidean distance in high-dimensional data does not faithfully reflect the similarity relationship of nonlinear samples.Therefore,a traditional t-SNE algorithm was proposed using Proposes the weighted Mahalanobis distance EMt-SNE based on information entropy.The experimental results show that: the data processed by EMt-SNE is more discriminative than traditional feature extraction algorithms,and thus the proposed method can greatly improve the performance of the classifier performance.(2)The traditional neural network training process has some disadvantages,such as parameter is difficult to select and complex network model.In this paper,gated recurrent unit(GRU)is proposed as the fault diagnosis model of steam compression refrigeration system.The diagnosis results are compared with the common fault diagnosis methods of refrigeration system: Extreme Learning Machine(ELM)and Support Vector Machine(SVM).The final results show that: for the Level1 fault level,the overall diagnosis accuracy of the proposed algorithm was 96.56%,which was best among the three classifiers.In addition,the accuracy of 4 types of local faults had reached 100%.Compared with the control group,the false alarm rate index decreased by 6.97%,5.99%,and 6.31%,and the false alarm rate decreased by 4.41%,0.64%,and 2.58%,respectively,verifying the effectiveness of the method in this paper.(3)In order to explore the actual effect of the proposed method,this article relies on the existing equipment in the laboratory and uses the C# language development tool to develop a data acquisition system for data acquisition.For a small air-cooled heat pump unit,common Fault simulation and data collection work has simulated three common faults,one of which is a concurrent fault.Based on the simulation results of the built diagnosis model and compared with the control diagnosis method,for the minor level fault Level1,the diagnosis accuracy rate of the EMt-SNE+GRU model is 97.5%,and the accuracy rate of the control group is 92.5% for ELM and 95.5% for SVM,GRU is 95.5%,and the results of the proposed methods are also better than those of the control group for the diagnosis of other fault levels.