Fault Diagnosis Of Centrifugal Chiller Based On Neural Network Method

The long-term operation of the chiller will inevitably lead to the performance degradation and failure.Preliminary study on chillers are mainly focused on the establishment of fault diagnosis model,improve the performance of the model,there is no research of chiller diagnostic performance under different fault severity levels,moreover,the chiller fault diagnosis model is less concerned with the data quality and feature information extraction.In this thesis,four fault diagnosis strategies are applied to the chiller fault diagnosis under different severity levels,which are verified by ASHRAE 1043-RP data.The neural network model is used to study the fault diagnosis of the chiller.The results show that the model can't diagnose the fault of the chiller due to the poor generalization ability of the neural network.In order to improve the fault diagnosis ability of the chiller model,the neural network is optimized by using Bayesian normalization.The optimized chiller model not only improves the average diagnostic accuracy of the chiller under different fault levels,but also can diagnose the excess oil and reduced condenser water flow fault with high accuracy.Although the optimized model has been improved from the view of a single fault diagnosis performance and average diagnostic accuracy,but when the chiller is in fault early,fault diagnosis efficiency is low for condenser fouling and refrigerant leak fault,and no more than 60%,furthermore,the average diagnostic accuracy of the model is low.In order to further improve the accuracy of the model,the author from the perspective of data quality,adopt the wavelet de-noising to improve the quality of data,eliminate data noise.The optimized data is used to build chiller fault diagnosis model,the results show that the accuracy of the chiller fault diagnosis model under different fault levels has been improved in some extent,especially for the first fault level,the average accuracy of the model is improved by 11%.The eight feature variables are manually selected to diagnose the seven chiller faults for the three fault diagnosis strategy,in order to extract the fault feature information intelligently,this thesis studied on the PCA(Principal component analysis)employed for feature extraction of chiller operation data to achieve the whole process of intelligent identify feature variable to fault classification.The results show that the number of principal components is proportional to the accuracy of fault diagnosis model,but PCA does not have the ability to remove noise,thus,in the case that the number of principal components and the number of characteristic variables are the same,compared with the fault diagnosis strategy based on wavelet de-noising and bias neural network model,diagnostic performance based on combination of PCA and bias neural network decreased slightly.In this thesis,four fault diagnosis strategies are proposed to improve the accuracy of the fault diagnosis model and the ability of intelligent extraction of the feature information.It has a certain application value and significance in the intelligent fault detection and diagnosis of chiller,which is worthy of further study.