A Study On On-line Condition Monitoring And Fault Diagnosis Of Yan'an Refinery's Flue Gas Turbine

Flue Gas Turbine is one of most common key equipment in petroleum chemical industry.The operation condition of Flue Turbine Set is related to the running period and energy consumption level of FCC directly,and it has important significance to ensure the production equipment run calmly,safely,effectively and energy saving.So the status monitor and fault diagnosis of flue turbine set could find the early latent fault of equipment,and according to fault development current could forecast fault,which could reduce grave accident and provide reliant technic and management guarantee for safe,steady,long-period and high-quality running.This article based on the research of the operation of Flue Gas Turbine,Cause of trouble,common fault diagnosis methods,the choice of characteristic signal and sensor,the current status and development trend of gas turbine vibration signal,achieve the fault sample data collection at the Distributed Control System(DCS)of Yan'an Refinery;Use the wavelet methods for the pre-processing of original fault data,accomplish 3 layer decomposition processing to the signal which is used after de-noising through adopting the db4 wavelet,extract the feature vector of the signal and normalization processing for it.Support vector regression is a theory of statistical regulation and learning methods which is specialized in studying the small samples conditions.It is well to solve the practical problems such as small sample,nonlinear and high dimension.Further discussion to the three important statistical theory that is VC dimension,empirical risk minimization and the structural risk minimization principle and structural risk minimization;Simultaneously,the SVM of linear,approximate linear,nonlinear and classification algorithm are studied,completing the study of SVM.Setting up the intelligent software platforms LIBSVM for fault diagnosis of Flue Gas Turbine,and combining with the signal that processed by Wavelet Packet,Used the pre-processing of feature vector as input sample,Adopted GA,Grid Search and PSO to parameters optimization in SVM,RBF kernel function and the best parameters <c,g> combination are selected in the finally for the sample classification.The results show that SVM in the Flue Gas Turbine system's fault diagnosis has a better recognition ability and feasibility which used PSO for parameters optimization.