Fault Diagnosis For Chiller System Based On Multi-source Information Fusion

The energy consumption of chillers accounts for the majority of the energy consumption in large buildings.Energy conservation research is an important way to build green-buildings.The deterioration of chiller's operating conditions and its faults may lead to much additional energy consumption,worsen quality of air in room and even reduce equipment's life.Therefore,it is extremely important for chiller to monitor online and diagnose faults.With the advent of the era of big data,a rapid development could be found about the technology of fault detection and diagnosis based on a mass of data.Many experts and scholars at home and abroad have carried out many extensive and profound studies,and achieved remarkable achievements.With the advancement of technology,some researchers have applied principal component analysis,expert rules,neural networks,deep learning,decision trees,support vector machines,etc.to the field of chiller's fault detection and diagnosis.The results of fault diagnosis using these algorithms are always in the form of boolean values.However,due to the complicated fault condition of the chiller and the insufficient measurement accuracy of the sensor,many uncertain factors appeared in the fault diagnosis process.If the fault diagnosis result is still presented with boolean value,it will cause a certain deviation.Thus,it is more reliable and reasonable to present the fault diagnosis result with probability value.Accordingly,Bayesian network is a network model combining graph theory and probability theory with the strong ability in uncertainty information processing.This thesis introduces the chiller's system and its typical faults and makes a thermodynamic analysis for typical faults.Furtherly,the Bayesian network model is built:1)determine the network structure based on expert knowledge and thermodynamic analysis of the faults;2)quantify the qualitative rules using the control limits of EWMA control chart,and count the frequency of residuals distributing in each threshold interval,and regard the frequency as a conditional probability between fault and fault symptom approximately.Additionally,considering the complexity of structure of Bayesian network model and many redundant variables included,this thesis proposes to integrate the sensitivity analysis into the Bayesian network and selects the most sensitive characteristic variables as the network nodes.At the same time,the incomplete problem for observed information is solved by merging non sensor information like uniting maintenance service records into the modle.Finally,the model was validated using the ASHRAE RP-1043 project data.The results show that the fault diagnosis results are good as a whole for the EWMA-BN model utilizing 19 nodes and 84 probability dependencies,but it is not as expected under lower fault level SL-1;the accuracy of diagnosis under all fault levels exceeds 0.98 and the most situations reach 1.0 for the EWMA-SA-BN model including 12 nodes and 17 probability dependencies.After merged the information such as historical maintenance record of the unit into the EWMA-SA-BN model,it still presents the good fault diagnosis performance when the observed information is incomplete.