Research Based On Data-driven Methods For Nonlinear Process Monitoring

With the continuous development of industrial systems toward large-scale,complex and highly integrated directions,requirements for safety and reliability of industrial systems are getting higher and higher.Effectively monitoring process in real time becomes an important topic.Sometimes modeling complex industrial systems is very difficult,even impossible,which makes model-based process monitoring methods are unavailable.Meanwhile,industrial processes can produce huge amounts of data,which implicit large amount of system information.Based on this phenomenon,it is meaningful to research data-driven process monitoring methods.Currently,a lot of research results are about linear algorithms for process monitoring,but these algorithms are not applicable to nonlinear processes.For this problem,two linear algorithms have been improved in order to applied to nonlinear systems in this paper.Firstly,the specific algorithm steps of PLS and T-PLS in industrial process monitoring are reviewed respectively.Since T-PLS algorithm is a linear algorithm and it is not suitable for monitoring nonlinear systems,kernel method is introduced into T-PLS to obtain the T-KPLS.Many outliers are often mixed in data which is collected from the actual industrial process.To tackle this problem,spherical method is introduced into T-PLS to obtain a new nonlinear approach,ST-KPLS.Using numerical simulation example,the monitoring ability of ST-KPLS algorithm is verified.Then,a modified algorithm for PLS,MPLS approach,is improved in this paper.Using MPLS,orthogonal partition for the original data space is realized,which makes the statistics and subspaces be completely matched.However,MPLS as a kind of linear method has limited ability to monitor nonlinear processes.In view of this problem,MKPLS algorithm is presented by combining kernel method with the MPLS and the monitoring capability of MKPLS is verified by numerical simulation.Finally,working principle and benchmark simulation of wastewater treatment process(WWTP)is introduced in the paper.Two kinds of abnormal events involved in WWTP is expounded and process variables and quality variables are reasonably selected.After that,the proposed two algorithms,ST-KPLS and MKPLS,are applied to monitor the abnormal events in the process to show their superiority...