Research On Diagnosis And Prediction Of Frosting Fault Based On Refrigeration System Of Cold Storage

The development of food cold chain can not be separated from the use of cold storage,but the energy consumption caused by cold storage can not be ignored either.The main energy-consuming equipment is the refrigerating unit,which will inevitably produce failures in the course of operation,it will not only affect food quality,increase system energy consumption,shorten equipment use cost,increase maintenance cost,and even affect the environment.Therefore,it is very important to predict and diagnose the faults of refrigeration system,to judge the types of faults in time and accurately,and even to predict the faults before they occur,in order to avoid the occurrence of faults.At present,the research on fault diagnosis and prediction in refrigeration field is mostly on air-conditioning,and the research on the refrigeration system used in lower temperature is less,and the research on the mode identification of evaporator frosting as the fault type is also less,in this paper,the Frost Fault of evaporator which is very common and difficult to avoid in cold storage is chosen as the main research content for pattern recognition and fault classification.At present,the methods for fault diagnosis are divided into analytic model,knowledge,signal processing and data mining.With the advent of big data,fault diagnosis can be done by data mining,which not only can update the fault diagnosis model according to the data,but also is relatively simple.The basis of deep learning in data mining algorithms is neural network,and neural network is not only used in data mining,but also widely used in artificial intelligence and pattern recognition,in this paper,the neural network algorithm is used to diagnose and predict the evaporator frosting fault in the refrigeration system of cold storage.Firstly,the thermodynamic analysis and the theoretical analysis of the main gradual type faults of the vapor compression refrigeration system are carried out,and the characteristic parameters needed by the system operation and the faults are selected for determining the measured parameters in the experiment,according to the reduction of refrigerating capacity,the experimental data of frost formation can be divided into three levels: light frost formation,medium frost formation and heavy frost formation.Then,BP Neural Network,Elman Neural Network,PCA-BP neural network and PCA-Elman neural network were established respectively from 127620 groups of frost experiment data under 18 different working conditions.The simulation shows that:(1)Reducing the target error can make the model more effective and the training accuracy higher.(2)Ten-fold cross-validation can improve the accuracy of BP neural network and Elman neural network,reduce the over-fitting,improve the model performance andgeneralization ability.The performance of BP neural network with and without 10-fold cross-validation,its optimal network accuracy is slightly higher than Elman neural network,but Elman neural network training time is shorter,faster.The comparison between BP neural network with and without 10-fold cross-validation and Elman neural network shows that Elman neural network with 10-fold cross-validation is better,and its training set,verification set and test set have higher accuracy,the training time of BP neural network with 10-fold cross-validation is less than 11 h.(3)The accuracy of the training set and the verification set of Elman Neural Network and BP neural network is similar under the same model parameters,but the accuracy of the test set of Elman neural network is more stable.(4)Compared with BP neural network,PCA-BP neural network can shorten the training time of the model while ensuring the accuracy.Compared with Elman neural network,PCA-Elman Neural Network has higher accuracy,that is,better generalization ability.(5)PCA-BP neural network and PCA-Elman neural network have little difference in the accuracy of training set and verification set,but under the optimal condition,the accuracy of PCA-Elman neural network test set is 93.167%,higher than that of PCA-BP network 91.279.Compared with Elman Neural Network,PCA-Elman Neural Network has higher accuracy of training set,verification set and test set,and the training time is similar,so PCA-Elman neural network has better performance.Therefore,the PCA-Elman model with the best number of hidden layer nodes can be used to get the best result after 10-fold cross validation.(6)The accuracy of all models can reach above 90%,and all models have good diagnostic performance.By comparing the results of PCA-Elman neural network with different training functions,it is better to select trainbr training function with shorter training time and higher generalization ability for 10-fold cross validation,the generalization ability of the optimal results is higher than that of the previous optimal results,and the training time is nearly one-tenth shorter than that of other 10-fold cross-validation networks.In contrast,the PCA-Elman neural network model established by trainbr training function can achieve better diagnosis effect under the optimal number of hidden layer nodes.