Complex Industrial Process Monitoring Methods Based On Local Information Mining

As modern industrial systems become increasingly large and complex,timely process monitoring and fault diagnosis technologies serve more and more important role to ensure process safety and improve product quality.With the application of distributed control systems,massive data are gathered,which leads that the multivariate statistical process monitoring(MSPM)methods have been one fascinating topic in process monitoring.However,the complexity of modern industrial processes does challenge the monitoring performance of traditional MSPM methods.Aiming at large-scale,nonlinear,non-Gaussian,and multimode characteristics of complex industrial processes,this thesis studies local information mining-based modified MSPM methods to content the practical monitoring demands.For large-scale process monitoring,a mutiblock PCA method based on variable subregion(VSR-MBPCA)is firstly proposed,where mutual information(MI)-based variable block division is performed and local statistics are extracted from the original PCA model.To obtain more sufficient local information,this thesis further proposes a novel mutiblock PCA method based on variable weight information(VWI-MBPCA).This monitoring approach divides the full components space into several subblocks and obtains the corresponding small-weight variable groups by investigating the weight information of variables upon the components,which can reflect local variable information more precisely.Simulation results on numerical example,Continuous stirred tank reactor(CSTR),and Tennessee Eastman(TE)process show that the proposed approaches can provide better monitoring performance than traditional PCA method.For monitoring nonlinear process with non-Gaussian characteristic and mining more fault information,this thesis proposes a modified KPCA-based process monitoring method based on double-weighted local outlier factor(DWLOF-KPCA).In order to avoid the assumption of Gaussian distribution,LOF method is introduced to construct monitoring statistics to monitor principal component subspace and residual component subspace in KPCA model,respectively.Further,to highlight fault information,a double-weighting strategy is designed to emphasize the fault kernel components and historical monitoring statistics,respectively.Simulation results on a numerical example and TE process demonstrate that this method can detect fault more effectively than traditional KPCA method.For monitoring process with linear-nonlinear mixture distribution,this thesis proposes a local variable characteristic analysis(LVCA)based hybrid modeling and fault detection method.Firstly,linearity and nonlinearity evaluating factors are defined to measure the correlation among different variables.Then,several linear and nonlinear variable blocks are divided based on the two factors,where PCA or KPCA models are built in different blocks and the Bayesian inference is adopted to build hybrid mode.Simulation results on a numerical example and TE process validate the efficiency of the proposed approach and show that the hybrid modeling strategy can adequately extract linear and nonlinear process information,which can remedy the weakness of single PCA or KPCA modeling.For monitoring multimode process,a double-level local information based local outlier factor(DLI-LOF)method is proposed in this work.Firstly,the local neighborhood standardization(LNS)strategy is employed to utilize the local data statistical information of mean and variance to handle the multimodality.Then,a novel variable LOF(LOF)is proposed to extract local variable information,which is further emphasized by a weighting strategy and the Bayesian inference.Simulation results on a numerical example and TE process verify the efficiency of DLI-LOF method for multimode process monitoring.