Research On Fault Identification Of Electronic Expansion Valve And Compressor Floodback Based On Deep Transfer Learning Model

The energy consumption of variable refrigerant flow systems has become an important part of the energy consumption of buildings in my country.It is very necessary to establish a fault identification model for the electronic expansion valve and compressor liquid floodback of variable refrigerant flow systems to reduce building energy waste and meet people’s requirements for increasing comfort and system operation stability.At the same time,in order to solve the problem of the small number of target samples,unbalanced data distribution and different distribution problems in the fault diagnosis process,which lead to the low generalization ability of traditional models,this thesis proposes a new method for fault diagnosis of variable refrigerant flow systems based on the deep transfer learning model.In this study,12 sets of experiments were carried out,and the electronic expansion valve fault data and the compressor normal operation and liquid floodback data were obtained to analyze the fault identification results of different distribution data and target sample data with less and unbalanced data.First of all,this thesis deploys the two-class fault recognition and multi-class fault recognition of variable refrigerant flow systems based on three commonly used deep transfer learning algorithms.They are the migration adaptive enhancement algorithm(TrAdaBoost),the deep neural network transfer learning algorithm(DNN-TL)and Joint Distributed Adaptation Algorithm(JDA).Then the key parameters of the deep transfer learning model are optimized.Finally,the model is analyzed from evaluation indicators such as accuracy,sensitivity,specificity,and hit rate,and it is combined with the supervised algorithm model,extreme gradient boosting(XGBoost),and deep neural network.(DNN)and convolutional neural network(CNN)for comparison.The results show that,on the one hand,when compared with the traditional supervised learning algorithm model,TrAdaBoost is 46%-51% more accurate than XGBoost.Compared with DNN,DNN-TL increases by 1%-4%.On the other hand,when three different methods of deep transfer learning algorithms compare the results of binary classification and multi-class fault recognition,TrAdaBoost is 2%-9% higher than DNN-TL,and TrAdaBoost is 38%-55% higher than JDA..The overall effect of TrAdaBoost model is slightly better than that of DNN-TL,which indicates that the deep transfer learning model has stronger generalization ability,and has a better diagnostic effect on data tasks with different distributions.Finally,as the size of the training data increases,the accuracy of DNN-TL is about 8% higher than that of DNN.For all training data sizes,especially in the small training data for electronic expansion valve fault identification,the accuracy of TrAdaBoost’s model is significantly higher.For DNN-TL,this proves the advantage of TrAdaBoost for transfer learning on small training data.In summary,the deep transfer learning model can significantly improve the fault recognition of variable refrigerant flow systems,and can improve the accuracy of fault recognition by learning similar data when the target data is less,and the data in the actual system is incomplete and incomplete.Data labeling provides a solution to practical problems such as manpower and material resources.With the upgrading of computer hardware equipment and the development of deep transfer learning technology,the deep transfer learning model proposed in this paper has certain research and application value in fault identification of variable refrigerant flow systems and related fields.