The Research On Fault Diagnosis Of Batch Process Based On Improved Neighborhood Preserving Embedding Algorithms

In order to improve production efficiency in modern industrial processes,the automation degree of production equipments improves continually,which makes the more and more complexity of production processes,so higher requirements for the safe and reliable operations of production processes are put forward.Because of its flexible production mode,batch production process can satisfy the rapidly changing market demand in modern society.It has received more and more attention.Therefore,how to establish a monitoring model of batch process for timely and accurate fault detection and diagnosis to ensure its production safety and product quality has become a problem which is to be solved in industry and academia.The accurate mathematical models of batch process are difficult to establish because of its complex characteristics.With the development of modern industrial technology,large amount of process data have been stored.To analyze and handle these process data can reflect the state of process operation.Therefore,multivariate statistical methods based on data have become important process monitoring methods.How to accurately and fully extract characteristic information which reflects the running status from massive amounts of process data is the key of process monitoring.Compared with PCA algorithm which focuses on global information,neighborhood preserving embedding(NPE)algorithm can obtain more characteristic information of the batch process because it excavates the local structural features of data.Aiming at the shortcomings of fault detection and diagnosis in batch processes,this thesis studies improved methods which are more suitable for batch processes based on NPE algorithm:(1)To solve the problem of insufficient feature extraction selected neighborhoods only by Euclidean distance of neighborhood preserving embedding(NPE)algorithm which causes bad effect of fault diagnosis,a novel fault diagnosis method is proposed by combining diffusion distance(DD)with the neighborhood preserving embedding(NPE)algorithm.Firstly,the intrinsic manifold embedded in raw high-dimensional data is fully discovered by reducing dimension of data.Then the features are extracted by learning the underlying geometry of the original data,and the local structure invariability is preserved in the data manifold,the problem of insufficient feature extraction selected neighborhoods only by Euclidean distance of NPE is avoided.Finally,T~2 and SPE statistics are used to detect faults,and the fault variable is diagnosed by the variable contribution graph method.The simulation results of penicillin fermentation process demonstrate that the proposed method is effective.(2)Aiming at the problem of fault detection caused by the dynamic characteristics of batch process data,a double weight multiple neighborhood preserving embedding(DWMNPE)algorithm is proposed.Firstly,by finding time neighbors for each sample point,the time correlation between samples is reflected.By defining angle neighbors,sample points are reconstructed by finding time neighbors,angle neighbors and distance neighbors for sample points.Three different manifold features can fully extract the essential structure of original data while the data dynamics are solved.In order to further prevent the loss of neighbor order information by NPE algorithm,a new objective function is constructed.The influence of the order information of three neighbors for the reconstruction error is considered while the minimum error is considered.Finally,the LOF statistic of the dimensionality reduction data is constructed to monitor batch process.The results of a numerical example and the penicillin fermentation process simulation demonstrate that DWMNPE algorithm for fault detection in dynamic batch process is effective.(3)Aiming at the problem of bad fault detection effect of batch process because of its non-linearity and mixed distribution of Gaussian and non-Gaussian,MDNPE-WDICA(multiwaydifferencialneighborhoodpreserving embedding-weighted and differencial independent component analysis)algorithm for fault detection of batch process is proposed.Firstly,the original data space is divided into Gaussian and non-Gaussian subspaces by using Jarque-Bera test method.Then,in Gaussian subspace,MDNPE algorithm is proposed by combining differential strategy with NPE algorithm to preserve the local structure invariant and deal with the nonlinearity of data while the dimension of data is reduced,this can overcome the computational complexity caused by the introduction of the kernel function.In non-Gaussian subspace,WDICA algorithm is proposed by combining weighted differential strategy with ICA algorithm to solve the nonlinearity of data while the data non-Gaussian information is fully extracted,and the local information of data is used effectively.Finally,a new monitoring statistic is established by Bayesian inference to realize fault detection for the whole batch process.The simulation results of penicillin production process demonstrate that the proposed algorithm is feasible and effective.