Research On Quality Related Fault Detection Method Based On Multivariate Statistics

With the integration and complexity of modern industrial systems,it is more and more difficult for model-based fault detection methods to establish accurate mathematical models for complex systems.Since modern industrial systems are equipped with a large number of sensors,the collection of data and work logs is more convenient and efficient.At the same time,data-driven fault detection technology has attracted more and more attention of scholars.Identifying whether process faults are related to product quality can improve product quality and extend equipment operating life by reducing unnecessary downtime and maintenance.The quality-related data-driven fault detection technology has attracted much attention in the performance monitoring of industrial systems.To solve this problem,this paper improves the deficiencies of PCA(Principle Component Analysis)and PLS(Partial Least Squares)in multivariate statistics,and verifies the effectiveness of the proposed method by simulation in classic numerical cases and Tennessee Eastman(TE)process.The specific work is as follows:(1)To solve the problem that PCA cannot detect whether the fault is quality-related,firstly,the relationship between process variables and quality variables is established by using principal Component Regression(PCR),and LU decomposition is introduced to further decomposition the process variable space into quality-related and quality-independent parts,and the validity of MPCR is verified in numerical cases.Secondly,aiming at the defects of the existing methods in detecting quality independent faults and regression faults in TE process,the data processing technology is used to screen the reactors highly related to quality variables and re model them,so as to reduce the interference of quality independent variables on the performance of the detection performance.Finally,in the Tennessee process,it is proved that the MPIPCR algorithm has better fault detection rate and regression fault detection ability than IPCR.(2)In order to solve the problem of false positives and missed positives caused by PLS's undesired diagonal decomposition of the process variable space,firstly,PCR was used to decompose the quality variable space and establish its relationship with the process variable space,and the process variable space was decomposed into highly related and irrelevant parts to quality.Secondly,in view of the problem of missed and false positives caused by the traditional fixed threshold that cannot consider the influence of recent samples,the idea of EWMA is introduced to set the adaptive threshold,and the effectiveness of the algorithm is verified in the linear and nonlinear classic cases and the TE process.