Fault Detection And Diagnosis Of Batch Process Based On Improved Gaussian Mixture Model Algorithm

Fault detection and diagnosis methods based on data-driven are widely used in complex industrial systems because they do not need to establish accurate mathematical models.Among them,the method based on multivariate statistics has always been a research hotspot in the field of industrial process monitorin g.Its core technology is that the method can extract useful information hidden in data fo r fault detection and diagnosis through the analysis and processing of industrial production process data.Based on summarizing the previous work and reading a large number of documents,this thesis proposes some corresponding improved methods for the problems and deficiencies that need to be solved in batch process monitoring based on Gauss Mixture Model(GMM)algorithm:1.Aiming at multiphase characteristic of batch process,the traditional methods often neglect the difference between different stages and do not consider the correlation characteristics and transition characteristics between stages,which affects the accuracy of process monitoring.In order to solve t his problem,a multiway adaptive Gaussian mixture model-kernel entropy component analysis(MAGMM-KECA)algorithm is proposed.Multiway adaptive Gaussian mixture model is used to automatically obtain Gaussian model information at different stages to realize multiphase flexible partitioning.The Gaussian model obtained is more accurate and more in line with the actual multi phase process.The different phases are modeled and monitored by using KECA.The CS(Cauchy-Schwarz)statistics are introduced to monitor the fault and the traditional contribution plot method is used to diagnose the fault.2.Aiming at the nonlinear and dynamic characteristics of batch process,a global-local regularized Gaussian mixture model(GLRGMM)algorithm is proposed.Combining neighborhood preserving embedding(NPE)algorithm to extract the local manifold structure hidden in high-dimensional data and principal component analysis(PCA)to extract the global structure,finally the global-local structure preserving is achieved,which is beneficial to maintain and acquire the nonlinear data structure to some extent.Then a new regularization term to GMM is defined to online monitor and update the Gaussian model to solve the problem of data dynamics.Finally,the integration of global-local monitoring indicators is used to effectively achieve online monitoring.3.Aiming at the nonlinear and multi-mode characteristics of batch process and the quality-related process monitoring,a multi way Gaussian mixture model-Concurrent Kernel entropy projection to latent structures(MGMM-CKEPLS)algorithm is proposed.At first,Renyi entropy is introduced into high-dimensional kernel space by using nonlinear projection.By calculating the contribution of Renyi entropy,the principal component is obtained,and the number of principal components is less,which reduces the computational complexity.Because the structural characteristics of different modal data are different,the modal da ta clustering of the multi-mode process is performed by GMM to obtain t he sample data of different mode.Then the respective model of Concurrent projection to latent structures(CPLS)is established in each mode.Finally,the integration unified monitoring statistic is built to achieve online monitoring and quality prediction by using the modal weight coefficient.4.Aiming at nonlinearity,multiphase,and mixture distribution of process variables in batch processes,a multiway weighted support vector data description algorithm based on similarity measure(Similarity Measure-MWSVDD,Sm MWSVDD)is proposed.First,aiming at the transition and timing correlation between different phases,a Gaussian mixture model of similarity measure is proposed to divide the multiphase process into stable phases and transition phases.By overcoming the shortcoming of traditional support vector data description(SVDD)in the construction of control limits,the sensitivity of SVDD to faults is improved.The D-test method is used to divide the distribution of variables into Gaussian distributions and non-Gaussian distributions.Then Gaussian and non-Gaussian variables are modeled and monitored by using Multiway Kernel Principal Component Analysis(MKPCA)and improved SVDD respectively at different stages.Bayesian inference is used to integrate a uniform amount of monitoring at each stage.Finally,the integration unified monitoring statistic is built at each phase by Bayesian inference.The data used in this thesis is generated by the standard platform of penicillin simulation experiment.Different initial conditions and fault parameters can be set manually.The simulation results show that the proposed algorithm s are effective by comparison with traditional algorithms.Finally,the main contents of this thesis are summarized and a number of topics worthy of further discussion and research are put forward.