Research And Application Of Data-driven Online Detection And Diagnosis Of TE Process Faults

Under the background of the increasing scale of industrial systems,the ways of interaction within the systems are becoming more and more diversified,and the complexity of the systems is increasing,resulting in more and more causes of system failures.As a result,there is great research value to use the massive historical data to detect and diagnose the industrial process fault,protect the normal operation of the system and reduce the loss caused by the fault.In this thesis,the Tennessee Eastman process is taken as the research background to carry out the research work of data-driven fault diagnosis method.There are three main problems in conventional data-driven fault diagnosis methods: Firstly,how to effectively preprocess the data for noise reduction and eliminate the influence of noise;Secondly,how to deal with nonlinear and distributed uncertain data;Finally,how to improve the traditional fault diagnosis method which only uses normal data modeling.In view of these problems,the main research contents of this thesis are as follows:(1)Aiming at the shortcomings that the conventional wavelet threshold de-noising method cannot determine the most suitable threshold function according to the characteristics of each layer of wavelet coefficients.This thesis improves the WT algorithm and proposes the AWT noise reduction algorithm to resolve these problems.AWT algorithm tests the normality of each layer of wavelet coefficients by introducing the normality test method,using the kurtosis divergence joint test method,and then selects the most suitable threshold function according to the normality of each layer of wavelet coefficients to achieve better noise reduction effect.(2)Aiming at the nonlinear and data distribution uncertain problems of the data generated by the Tennessee Eastman process,the optimal kernel entropy vector analysis method OKECA is introduced.This method can not only deal with the Non Gaussian multivariate data,but also deal with the nonlinear data.(3)Aiming at the defect that the conventional OKECA fault diagnosis algorithm does not use fault data modeling with known fault types,this thesis improves the OKECA algorithm and proposes FDE-OKECA algorithm.FDE-OKECA constructs the normal data sub-model and extracts the normal data kernel entropy component by adopting OKECA method,constructs the fault data sub-model and extracts the fault data kernel entropy component by adopting OKECA-LDA method,converts the statistics of the sub-model into the fault probabili by adopting Bayesian inference,and finally calculates the overall statistics weighted by the probabilities of the two sub-models,thus effectively enhancing the fault diagnosis capability of the algorithm.In addition,the new algorithm uses fault data to build an OKECA fault classification model to judge the fault type.When the model detects a fault,the OKECA fault type discrimination model is activated to determine the fault type.(4)In this thesis,a new fault diagnosis strategy is proposed by combining AWT with FDE-OKECA,and it is applied to the TE process.The new strategy first uses the AWT method to de-noise the data,then uses the FDE-OKECA method to diagnose its fault,and finally uses the TE process simulation platform to verify the performance and necessity of the proposed algorithm.