Research On Methods And Application Of Fault Diagnosis For Turbo-generator Unit Based On Bayesian Network

With the gradual adjustment of China's energy structure,the installation of renewable energy keeps growing.Large-capacity turbo-generator units have to participate to peak-shaving operation and operate under variable conditions.In addition,power electronic devices,such as series capacitance compensation and high voltage direct current transmission,are widely used in power grid,which makes the structure of power grid more and more complicated.Owing to the circumstance of electro-mechanical coupling,the risk of faults has increased.Based on the fault mechanism research,signal analysis,artificial intelligence and other methods were combined in this thesis,realizing the fault diagnosis for turbo-generator units,which is of great significance to ensure the safety and efficiency operation of the units.Firstly,the formation and evolution laws of complex equipment faults were analyzed,based on system theory and faults characteristics.Fault mode and effect analysis(FMEA),fault tree analysis(FTA)technologies and domain knowledge involved in the operation of turbo-generator units were used to analyze the cause-effect chain,fault effects,treatment measures and fault features of the 11 types of typical faults.And then,the multiple diagnosis information was obtained,which is a important part of big data for diagnosis.Big data involoved in fault diagnosis should take into account both the quantity and the comprehensive of data.Secondly,based on the fault mechanism anlysis results,fault features extraction methods were proposed.The vibration frequency domain features and torsional vibration damping characteristics were extracted by the spectrum analysis and enveloping fitting technology.Considering various working conditions of turbo-generator units,signal analysis and artificial intelligence technologies,such as kernel density estimation and support vector regression,were applied to determine the reference interval and threshold feautres extraction to decouple the features from the working conditions which can detect the abnormality earlier.In additon,qualitative trend analysis and correlation coefficient method were adopted to identify the time-trend and correlative features,after which,the recognition rules were given.Fault feature extraction requirements were made from the application side to achieve the goal of fault diagnosis.The extracted features can express equipment status more effectively and better identify the fault type.Thirdly,poor diagnosis results were obtained because of only using a small amount of information,such as vibration spectrum.The Bayesian network(BN)models for fault diagnosis of turbo-generator units were established by integrating FTA and FMEA results in order to integrate multiple information effectively.More scientific results could be obtained by means of the BN models,which comprehensively consider the knowledge of fault causal chain and fault features.The reasonable and objective assignments of BN parameters were performed by combining different approaches,such as the logical relationship between the faults and features,the independent model of causal effect and so on.The joint tree algorithm was used for the diagnosis and reasoning.Thus,the certain faults,the excluded faults and the possible faults were obtained.What's more,the probabilities of faults were sorted,so that the diagnosis results could be more effective to guide operations and maintenance.Finally,the condition monitoring and fault diagnosis system of the turbo-generator unit was designed and developed based on the multi-agent technology for the health management of the unit.The function modules of data collection,data analysis,fault diagnosis and health evaluation were shown in this thesis.The monitoring and diagnosis prototype system of 1000MW turbo-generator units has been successfully developed by use of the advanced data acquisition,data communication,programming development and large-scaled relational database technologies.