Research Of Complexed Process Fault Detection Based On Neighborhood Preserving Embedding

With the development of modern industry towards complexity,intelligence and digitalization,the requirements for product quality,production safety and reliable operation are constantly rising.It is a challenge of modern industry to ensure the production safety and improve product quality.Due to the extensive use of distributed control systems and the rapid development of computer technology,mass data reflecting process operation status and product quality have been collected and stored.Therefore,process monitoring methods based on multivariate statistical process monitoring have been rapidly developed.Traditional multivariate statistical methods assume that the process is not affected by noise and outliers and has characteristics of static,uniform conditions and linear relationships between variables.However,with the complexity and large-scale of industrial processes,process data don't satisfy the above assumptions.Aiming at the characteristics of large scale,multi-operation unit,nonlinearity,dynamics and multi-distribution of complex processes,as well as the problems of fault feature enhancement and noise suppression,based on neighborhood preserving embedding algorithm,this dissertation analyzes these process characteristics and proposes improved algorithms for process fault detection combined with information extraction strategy.The main research contents are as follows:(1)Because neighborhood preserving embedding(NPE)algorithm is easily affected by noise and outlier,and cannot consider the neighbor element distance when preserves the local structure.A fault detection algorithm based on sparse weighted neighborhood preserving embedding(SWNPE)is proposed.The optimal sparse representation is obtained in the neighborhood,and the noise and outliers are removed while the computational difficulties caused by the global optimization are avoided.Since the influence of the neighbor element distance is not considered in obtaining the neighborhood,nearer neighbors are more important for the local structural preserving than farther neighbors.Therefore,according to the neighbor element distance,different weight values are given when the optimal sparse representation is calculated to fully extract the neighbor structure,and an enhanced objective function is established to obtain local sparse structure.A numerical process and Tennessee Eastman(TE)test platform verify the fault detection efficiency of the proposed algorithm in large-scale complex process.(2)A nonlinear dynamic process monitoring algorithm named sparse representation preserving embedding extreme learning machine(SRPE-ELM)is proposed for nonlinear and dynamic characteristics of complex process.The algorithm inherits the ability of unsupervised extreme learning machine which can fast extract nonlinear manifold structure of industrial process.The sparse representation preserving embedding is applied to remove the noise and construct the adjacency graph with data self-adaptive neighborhood.SRPE-ELM can avoid the selection of neighborhood parameters and self-adaptively extract the dynamic manifold structure of process data.A numerical process and TE test platform verify the fault detection efficiency of the proposed algorithm in nonlinear dynamic process.(3)In order to fully extract the global and local features of nonlinear dynamic batch process,a multiway dynamic nonlinear global neighborhood preserving embedding(MDNGNPE)algorithm is proposed.First,the time-lagged window is used to remove the auto-correlation in time series of process variables.Secondly,a polynomial mapping algorithm is constructed to remove nonlinearity of process variables,avoid unnecessary redundancy and reduce the computational complexity.Thirdly,the global and local structures are preserved by using global neighborhood preserving embedding algorithm.Different from the basic kernel mapping method of nonlinear mapping,MDNGNPE algorithm considers the nonlinear characteristic with many physical constraints,and preserves the global and local data structures in dimensionality reduction concurrently.A numerical process and the penicillin fermentation process verify the fault detection efficiency of the proposed algorithm in nonlinear dynamic batch process.(4)Aiming at multi-phase and the feature extraction problems of batch process,a multi-phase batch process monitoring algorithm based on multiway weighted global neighborhood preserving embedding(MWGNPE)is proposed.First,for the multi-phase feature of batch process,gaussian mixture model is used to divide phases by clustering characteristic.Secondly,after the multi-phases have been divided,global and local structures are extracted by using GNPE.Thirdly,probability density estimation characteristic of gaussian mixture model is introduced to estimate the probability density of the extracted global and local structures,and then construct weighted matrix to enhance useful information and suppress noise.This algorithm can effectively capture the fault information which hidden in the process data.The effectiveness and advantages of proposed method are verified by a numerical system and the penicillin fermentation process.The results show that the proposed algorithm has the superiority in multi-phase batch process monitoring.(5)For the related or independence relationships between variables of batch process.A batch process monitoring algorithm based on weighted global neighborhood preserving embedding(WGNPE)and greedy support vector data description(GSVDD)is proposed.First,the related variables and independent variables are separated by mutual information.Secondly,WGNPE algorithm is used to extract the global and local structures of the related variables in batch process and highlight the fault information.Thirdly,GSVDD algorithm is used to extract the process information of independent variables quickly and effectively.Finally,statistical models are established in the related variables space and independent variables space to realize process fault detection.The traditional monitoring algorithms usually carry out a single statistical model according to the related or independent relationship,and the fault feature information is not fully taken into account.Therefore,this algorithm establishes a corresponding statistical model for fault detection according to the related and independent relationships of process data.Maintaining the advantages of WGNPE and GSVDD algorithms while removing the interference from unrelated variables.The effectiveness of the proposed method is verified by the penicillin fermentation process.