Research On Data-driven Fault Diagnosis Method Based On Multivariate Statistical Analysis

In recent years,the process industry has been developing towards integration and complexity,and effective fault diagnosis technology is the key to ensure system safety and productivity,and therefore it is also receiving more and more attention.Data-based methods do not rely on process knowledge and can achieve reliable process monitoring,which has attracted widespread attention.Due to the complex structure of the data,there are problems of non-Gaussian,non-linear,multi-modal,etc.,resulting in fault classifi-cation results that fail to meet expectations.Therefore,this paper addresses the theory of data-driven fault diagnosis based on statistical methods and the main points are as follows:(1)For the characteristics of complex industrial processes,a fault diagnosis method based on variable-weighted separability-oriented subclass discriminant analysis(VW-SSDA)is proposed to solve the problem of non-Gaussian,multimodal,and overlapping process data in the projection space.The variables are weighted based on their correlation with faults? clustering based on separability criteria is used to divide each type of fault data into subclasses,which are used to redefine and optimize the FDA scatter matrix.Through a large number of experiments targeting the TEP process,it is verified that the method can retain more classification information and significantly improve the accuracy of fault classification.(2)A local class-specific discriminant analysis with variable weighting(VW-LCSDA)model is proposed to address the problem of over-dimensioning of FDA-based fault di-agnosis methods and ignoring local structural information of data.The variables are weighted by assessing the impact of the fault on each feature.Besides,feature extrac-tion and classification were performed using an in-class and out-class scattering matrix of CSDA(class-specific discriminant analysis).In addition,the local structure of the sam-ple data is preserved with reference to the LPP(Locality preserving projection)method,which is used to optimize the scatter matrix.The performance of this method has also been verified experimentally.(3)Two nonlinear fault diagnosis methods based on kernel separability-oriented sub-class discriminant analysis(KSSDA)and kernel local class-specific discriminant analysis(KLCSDA)are proposed for the nonlinear characteristics of process data in the process industry.High-dimensional mapping of nonlinear data is performed utilising the kernel method,and linear relationships of the data are found in high-dimensional space using the previously mentioned SSDA,LCSDA methods.The effectiveness of the method is verified by experiments based on the TEP process.(4)A fault diagnosis method based on parallel Adaboost is proposed.In view of the stability and generalization of the ensemble method,Adaboost is used for fault diagnosis,and a weighted fusion approach is used to construct parallel models.Before training the parallel model,new training datasets are generated by evaluating the correlation between variables and faults and the correlation are utilised as a probability for feature extraction.Applying FDA as the base classifier.The validity of the model is verified by experiments.