Fault Detection And Diagnosis For Data Center Air Conditioning System

The energy consumption of data center is really huge,while 40% of it is composed by air-conditioning system.Applying fault detection and diagnosis technology on air-conditioning system,which conforms to the trend of energy-saving and emission-reduction,could effectively reduce the malfunction of air-conditioning and reducing the energy consumption.Malfunction of air-conditioning system can be divided into two different types.One is component fault,the other is sensor fault.In this thesis,both types were studied and effective FDD method combined with deep learning and others technologies were proposed.Because the two types of malfunction are not isolated,and the interact between the two types adds to the difficult of diagnose,studies on the hybrid of two kinds of malfunction are necessary.In sensor t fault diagnosis,this paper proposes a method based on LSTM,which identifies fault sensor by time series analysis.The FDD models were built for liquid line temperature sensor and discharge temperature sensor,and succeed to identify the fixed biases and drifting biases of the two sensors.In addition,the cost,accuracy and expansibility of LSTM method and hardware redundancy method were compared,respectively,which leads to the conclusion that LSTM has practical application value.In the component fault diagnosis,refrigerant undercharge and EEV malfunction were studied.A model-based method was applied,fault-free models were established by multiple regression analysis.Then the fault diagnosis was implemented based on the residual between the output of the model and the actual measured value.The method is simply to implement,but accuracy of this method is varied with different fitting objects according to test results.Due to the weakness of this method,a method based on random forest was proposed.The diagnostic accuracy and detection accuracy of this method are all above 99%.To prove the practicability of this method,cross-failure testing and cross-condition testing,which simulate the actual situation,are also carried out.In the last,the influence of the faults between component fault and sensor fault were analyzed.It is found that fault diagnose under hybrid of two kinds of faults requires the both result of sensor FDD model and component FDD model.Fault location under hybrid faults can use rules,which describe the relationship between different model diagnosis result and specific fault type.The rule of manual extraction is not practical enough because of its lack of detail.Therefore,a mixed fault diagnosis method based on rules-random forest was proposed and tested.The results show that the method can accurately identify the specific fault type under hybrid situation.