Fault Detection And Diagnosis Of Screw Chiller Based On Diagnostic Bayesian Network

Due to the increasing complexity of chiller system and the incomplete observation of the information,mappings between various faults and their corresponding symptoms are uncertain.Therefore,the probability method can be used to solve such problems well.In this paper,seven kinds experiments are carried out on screw chiller under different working conditions.Based on the experimental data,the fault detection and diagnosis method of chiller using Bayesian network theory is studied.First,based on the analysis and summary of the existing research results,and combined with the existing experimental conditions,5 kinds of independent faults and 2 kinds of mixed faults were selected as main research objects,and 7 kinds of fault experiments were simulated to obtain the data basis.The fault detection and diagnosis model is built by using diagnostic Bayesian network(DBN)through analyzing the fault and its symptoms.In order to solve the problem of fuzzy fault-free boundary caused by fault detection and diagnosis synchronously in Bayesian network,a support vector data description algorithm(SVDD)model is proposed to separate fault detection process from fault diagnosis,so as to determine the learning boundary of fault-free condition.The simulated annealing algorithm(SA)was used to determine the super-parameters of the model.The validity of the model was verified by using the experimental data.And for the fault that could not be separated effectively with similar symptoms,the method of separating diagnosis by adding field operation and integrating operational evidence was proposed.Based on the method of adding operational evidence,the model structure was further optimized for the two mixed faults including thermal and sensor faults.By analyzing the variation of cooling capacity loss of the chiller under different faults,the mapping relationship between each fault and the new evidence was matched for the diagnosis of mixed faults.The method of fault diagnosis is compared with the method of using the residuals of the non-fault model directly.The experimental results show that the proposed model can effectively diagnose the fault.The fault detection accuracy reaches 93.6%.Diagnosing rate of refrigerant leak and TCO sensor bias reaches 100%,decreasing of chilled water flow rate reaches77.8%.6 samples from TEO sensor bias and decreasing of frozen water flow rate conditions are all successfully separated.Moreover,two kinds of mixed fault diagnosis rate reach 75% and 41.6% respectively.