Fault Diagnosis For Chiller System Based On Bayesian Network And Its Optimization By Association Rules Mining

During recent years energy conservation is already the main guideline of refrigerant industry.And most of energy wasting comes from system faults.As to such a problem,a fault diagnosis system based on bayesian network is proposed.After evaluation the network is proved to be effective to diagnose faults.In this study bayesian diagnosis network which stimulates the thinking mode of technicians and experts is constructed for chiller system.The structure and probability distributions are two significant elements of bayesian network.The network structure of chiller system is divided into three layers: additional information layer,fault layer and fault symptom layer.Its probability distributions are maily obtained by statistic analysis and experts' knowledge.Once constructed,the posterior probabilities can be calculated while some evidences are input into the network.And the fault can be diagnosed and isolated by comparing the posterior probabilities of different faults.After fault isolation,the bayesian diagnosis network is validated by data sets.In the process performance data are divided into train set and test set.Two thirds of data,as the train set,are used to obtain the structure and probability distributions;One third of data,as the test set,are used to evaluate the effective of the network.After evaluation,the accuracy of bayesian network can be up to 90% especially for some typical faultsAs for chiller faults that has not shown good evaluation results,association rules mining is used to optimize its network structure and parameters.After optimization,the fault diagnosis results also need to evaluate.The result shows that the optimization has a good impact on the diagnosis accuracy of non-condensable gas fault which has increased by 16%.Unlike traditional Booleans probabilities are applied to present the uncertainties in measurements,experts' knowledge,fault patterns and symptoms.Meanwhile due to uncertainties,one piece of fault can results in multiple fault symptoms and different faults also may lead to same symptoms.As a consequence it is more reasonable to present the faults diagnosis results with probabilities rather than Booleans.Besides the maintance records and service history can also be used to diagnose faults except sensor measurements.In theory,the more information the network uses,the more precise the network is.In summary,bayesian network is suitable for chiller fault diagnosis from theory and application.