Research On High Performance Fault Diagnosis And Health Assessment System For Chillers

As the key equipment of the refrigeration system of data center,the main function of the chiller is to provide the cold source for the data center,ensuring the environmental quality of the machine room,and the reliability and safety of the equipment in the machine room.In addition,the faults of the chillers in actual operation are mainly gradual faults,which will take a long time from occurrence to obvious fault symptoms.This kind of operation with faults will additionally increase the energy consumption of the refrigeration system.Aiming at the above problems,this paper designed and implemented a high-performance fault diagnosis and health assessment system for the chillers.The main work and results of this study are as follows:Firstly,a high performance fault diagnosis method based on one-dimensional convolution neural network and gated recurrent unit is proposed to solve the problem of general feature extraction ability and poor generalization ability of the existing chiller fault diagnosis methods.This method combined the speed of one-dimensional convolutional neural network with the high precision of the gated recurrent unit.First,the local features of faults are extracted by training one-dimensional convolutional neural network with the randomly ordered historical data;then,the parameters trained by one-dimensional convolutional neural network are applied to the hybrid neural network model(1D-CNN＿GRU)proposed in this paper,so as to retain the long-term dependency between data.The experimental results show that,compared with other mainstream methods,1D-CNN＿GRU not only has good generalization ability,but also has better performance on the public data set RP-1043 and the real data set of the data center.The diagnosis accuracy of different fault levels has reached more than 90%.Secondly,in order to solve the problems of the difficulty of feature extraction and the high complexity of quantization algorithm in the process of health evaluation of the chillers,a health evaluation method based on the combination of LSTM-autoencoder and euclidean distance was proposed.Aiming at the difficulty of fault feature extraction,this research only adopted the normal state feature data to unsupervised train the LSTMautoencoder,and constructed the normal sample feature space.The euclidean distance was introduced to measure the degradation distance between the real monitoring vector and the feature space,and to measure the degradation degree of the chiller,which solving the problem of high complexity of the health measurement algorithm.The experimental results show that the health degree calculated by the method proposed in this study is consistent with the degradation degree of the fault,and it has better industrial quantitative evaluation value.Finally,based on the above research results,the health management system of the chillers based on Spring Boot framework was designed and implemented.The system takes the real-time sensor data of the data center as the data source,and passes it into the database through its internal BIM interface for structural transformation.The system displays the status monitoring results and model training results by calling Echart open source visualization library.In addition,the system also includes code generation,form management and other modules,which not only reduces the workload of repeated code writing,but also provides technical support for operation and maintenance personnel,thereby saving maintenance costs.The realization of this topic is of great research value and practical significance for maintaining the normal environment of the machine room,troubleshooting early faults,reducing equipment loss and creating "green data center".