Research On Refrigerant Charge Abnormal Diagnosis Of VRF System By Deep Auto-encoding Model

The control strategy of VRF systems is complicated,and equipment failure is prone to occur after long-term operation,which increases equipment energy consumption.The data-driven technology is more effective and convenient in application because it does not require downtime maintenance and professional personnel.This thesis takes the data-driven model as the entry point,and takes the VRF refrigerant charge failure as the research object,uses deep auto-encoder research on the characteristics of VRF data,the diagnosis effect and the generalization ability of the model.In this thesis,the following work has been carried out.First of all,this thesis researches and analyzes the VRF operation data of the enthalpy difference laboratory,collects the operation data of the cooling and heating operation mode under different working conditions in the experiment,and then classifies them.It includes a total of nine refrigerant charge levels ranging from 63.43% to 130%.After preliminary analysis of the data distribution of different labels after dimensionality reduction,it is found that the deep auto-encoding model can reduce the data dimensionality while reducing the information loss of the original data as much as possible while reducing the data dimensionality.When the original VRF data is reduced to 6 dimensions,the information loss is only 0.067,which is far better than the information loss of PCA(0.394).In this thesis,on the basis of obtaining dimensionality reduction data,diagnostic accuracy and F1 score are used to evaluate the BP classification results of these dimensionality reduction data.According to the characteristics of the deep auto-encoding model,model fine-tuning and grid search technology are introduced to optimize the model and parameters,and finally a relatively excellent diagnostic model is obtained.After adjusting the model and parameters,the diagnostic performance of the DAE-BP model has been significantly improved,and the diagnostic accuracy has increased from 71.4% to96.8%.Finally,in order to further analyze the generalization ability of the model,this paper uses data samples in different operating modes as training sets to test the diagnostic performance of data in other operating modes,and try to optimize from the data side and the model side.The results show that there is a big difference in data between cooling and heating operation modes,which will lead to a significant decrease in diagnostic performance,and this situation cannot be optimized by adding a Dropout layer.However,when the operating data in the cooling and heating modes are used as the training set at the same time to construct the DAE-BP diagnostic model,the diagnostic effect of the model in both operating modes exceeds 92%.This situation also appears in the MIC-BP and PCA-BP models.In response to this phenomenon,this thesis proposes a VRF online data circulation scheme based on blockchain technology.Research shows that this solution can enable VRF operating data to break through the trust gap between individuals and enterprises,and realize the circulation and storage of absolutely true,completely credible and non-tamperable VRF key data.And because there is no transaction data,the storage cost is much smaller than that of the traditional blockchain platform,and the storage cost of a single multi-online data group is less than 2kb.