Study On Key Performance Indicator Related Fault Detection Approaches Based On PLS

With the safety and reliability of industrial production systems continuously improving,data driven process monitoring methods have become a hot topic in the academic field.It can detect the faults in the industrial process quickly and accurately.And it ensures the stability of the system by handling the fault.However,research reports and industry feedback show that not all process failures affect product quality.If some unnecessary alarm are neglected,the maintenance time and manpower can be reduced,and the production efficiency can be improved..Therefore,a new type of fault detection method arises at the historic moment,that is,key performance indicators(KPI)related fault detection method.The method can effectively distinguish whether the occurred fault affects the KPI in the industrial process.However,the method can't solve the problem that the data is mixed with outliers.This paper modifies the partial least squares(PLS)in multivariate statistical analysis technology,and integrates it with KPI fault detection technology.For linear and nonlinear static systems,robust KPI-related fault detection methods are studied.The main research contents are as follows:(1)In the linear process,in order to solve the problem that PLS is sensitive to outliers and the complexity of its fault decision logic,this paper introduces a robust PLS algorithm to reduce the influence of outliers on the model.After that,the singular value decomposition(SVD)is used to decompose the process variable space into two subspaces,KPI-related subspace and KPI-unrelated subspace.Thus,a robust KPI prediction and linear fault detection method is proposed.It can distinguish whether the fault has affected KPI.The two algorithms perform KPI prediction and KPI related fault diagnosis with TE process.To compared with PLS,the proposed method has stronger robustness and better fault detection performance.(2)In this paper,an modified spherical KPLS(MSKPLS)on-line detection method is proposed against outliers in nonlinear process.In this method,the high dimensional feature variables obtained from the kernel processing are projected onto a spherical surface with spherical strategy.Then,performs KPLS algorithm with gained data.The coefficient matrix obtained by modeling the variable space is decomposed by SVD to get two orthogonal subspace.The fault can be effectively monitored.This method not only can reduce the influence of outliers on the modeling accuracy,but also can solve two key problems,i.e.nonlinear problem and KPI related detection.Numerical examples and TE chemical systems are used to compare the existing algorithms with the new method.The results show that the new method has better performance in KPI-related fault detection.