Nonlinear Process Monitoring And Fault Identification Based On Adaptive Moving Window Kernel Principal Component Analysis And Kernel Density Estimation

With the increasing complexity of the contemporary chemical production process,the increasing demand for social productivity,the growing scale of production,and the increasing popularity of automated machinery production,a complete process monitoring system can detect faults in the production process in time and provide safe and reliable technology for enterprise development.Guarantee,while reducing the waste of raw materials and improving product revenue,it has become a necessary monitoring system for most industrial production.In nonlinear process monitoring,many process faults are difficult to monitor and identify.In this paper,multivariate statistics and kernel density estimation methods are used as the core to study industrial chemical process monitoring and fault identification.Principal Component Analysis（PCA）is a relatively basic fault monitoring method in the multivariate statistical process.This method does not do much processing on the original mathematical model,and only needs to use the projected data through the mapping function.Failure analysis.Although PCA is widely used in fault monitoring research,it can only be limited to fault monitoring in linear systems,and is obviously not suitable for dealing with faults existing in nonlinear systems.Since kernel learning has strong generalization ability,many scholars have introduced it into nonlinear process fault monitoring.Kernel Principal Components Analysis（KPCA）is the most commonly used method in nonlinear fault diagnosis and identification technology.Dimensional space performs principal component analysis on the data.This method saves the complicated computation of other monitoring methods,and only needs to deal with eigenvalues in the high-dimensional feature space.The selection of kernel parameters in KPCA seriously affects the monitoring performance of process data.Advanced optimization algorithms such as particle swarm optimization can be used to ideally solve the problem of kernel parameter selection in KPCA.However,when there is a small disturbance in the nonlinear process,the monitoring effect of KPCA will be untimely or missed.Aiming at this situation,an Adaptive Kernel Principal Components Analysis（AKPCA）method is proposed.In order to further shorten the monitoring time and improve the performance of fault detection rate,a nonlinear process fault monitoring and identification method（Adaptive Kernel Principal Components Analysis-Kernel Density Estimation,AKPCA-KDE）combines AKPCA and Kernel Density Estimation（KDE）.With the help of Tennessee-Eastman process platform（Tennessee Eastman,TE）as a simulation experiment platform,the monitoring performance such as fault diagnosis rate,false alarm rate and missed alarm rate were comprehensively compared and analyzed.The main research contents and conclusions are as follows:1.This paper introduces the monitoring methods of PCA and KPCA successively,and takes the fault of the TE simulation platform as an example to conduct a comparative experiment.There are still many false positives in PCA.From the fault monitoring diagram of KPCA,it can be seen that KPCA has no false positives.In this case,the KPCA method has obvious advantages over the PCA method in dealing with nonlinear problems.2.Combining the basic principles of PCA,KPCA and variable moving window,this paper proposes a moving window AKPCA monitoring method to reduce the false alarm rate of the monitoring model due to slow process changes in the real chemical environment.The basic idea of the method is to first employ a multivariate exponential moving average（MEMA）to predict the process mean shift,and then combine the estimated mean shift with the components extracted by KPCA to construct an adaptive monitoring statistic.Simulation experiments show that by updating the size of the moving window in each step,the scaling parameters,forgetting factor,control limit and monitoring model can be updated at the same time,and the problem that traditional PCA is easy to lose important information in the process of time-varying monitoring can be further avoided.3.Aiming at the problem that the control limits（CLS）of Gaussian distribution in traditional KPCA will reduce its monitoring performance,this paper uses the kernel density estimation method to determine the control limits of the moving window AKPCA,which combines the moving window AKPCA with the kernel density estimation（Adaptive Kernel Principal Components Analysis-Kernel Density Estimation,AKPCA-KDE）is used for nonlinear process fault monitoring and identification.This method is applied to TE process simulation monitoring,and the detection rate of AKPCA-KDE for all 21 faults is found.Compared with KPCA’s control limit method using Gaussian distribution,AKPCA-KDE has a higher fault detection rate;The detection latency of the method is equal to or less than that of the other methods.By changing the bandwidth of KDE and the number of retained pivots for fault detection and identification,the traditional KPCA recorded a higher false alarm rate（FAR）value,while the AKPCA-KDE method did not record any false alarms.It is obvious that AKPCA-KDE has better monitoring performance and provides a new monitoring and fault identification method for industrial process monitoring.