Research On Modeling And Monitoring Of Complex Industrial Processes Based On Non-gaussian Feature Extraction

Process monitoring technology can ensure the safety of production process,among which the multivariate statistical process monitoring(MSPM)is an important research object in the current process monitoring direction.The traditional MSPM has too many restrictions,and the data of complex industrial processes usually do not conform to these assumptions.It cannot be completely described by the Gaussian distribution model alone.Therefore,it is necessary to extract the non-Gaussian feature in the process and the process is monitored based on the extracted non-Gaussian feature.In addition to non-Gaussian feature,process data also has problems such as multimodal,large-scale,transition process and outliers,so it is necessary to analyze the relevant characteristics one by one based on the non-Gaussian extraction method for fault detection.(1)Aiming at the steady-state conditions under the non-Gaussian influence of process data,a fault monitoring method for industrial production processes based on the double-layer non-Gaussian monitoring(DLNGM)algorithm was proposed.NonGaussian and Gaussian double-layer ideas are introduced.Different models are used for high-order information and low-order information for fault monitoring.This method fully considers the complex distribution of variables and solves the difficult problem of non-Gaussian process fault detection.(2)For large-scale distributed processes and situations where information for partitioning molecules is often difficult to obtain.A double-layer non-Gaussian monitoring fault detection method based on hierarchical multi-block decomposition(MB-DLNGM)is proposed.Introduce mutual information and double-layer nonGaussian monitoring methods.The process variables are divided into multiple subblocks with mutual information.For each sub-block,a two-layer non-Gaussian monitoring model is established,and non-Gaussian and Gaussian monitoring statistics are constructed to achieve the purpose of fault detection.(3)Aiming at the problems of processes of identification and monitoring in timevarying and multi-modal situations and simultaneously considering the complex distribution of data,we proposed a fault detection method based on sample-based multi-modal classification and hierarchical non-Gaussian algorithm.The training data are projected to the DLNGM model by means of loop iteration,and then the corresponding prediction residuals are calculated using the DLNGM algorithm.Based on these residuals,the multimodal process data are classified from the sample dimensions.Afterwards,a corresponding DLNGM model is established in each of the classified modalities,and finally online recognition and fault monitoring are performed.(4)Aiming at the problem that outliers and non-Gaussian characteristics can exist in the process data at the same time,we extended the modeling method based on the global similarity of student's-t mixture model distribution to transition process fault monitoring.The student's-t distribution mixture model can not only deal with nonGaussian characteristics,but also reduce the influence of outliers on modeling,which greatly improves the robustness.