| In modern industrial production,with the large-scale application of sensors,PLC and DCS systems,data processing technologies have attracted much attention.The data-driven multivariate statistical process monitoring method can online evaluate the process operation status online,thus ensure the reliability and stability of the system,which means significantly for the enhancement of the quantity and quality of the products.However,outliers and missing data commonly existing in real industries challenge the data-driven process monitoring methods.In this dissertation,the problems of process monitoring are studied based on the modified projection to latent structures(MPLS)model.The details are listed as follows:(1)Considering that the process data contains outliers,this dissertation analyzes the robustness of the MPLS model and proposes a robust MPLS algorithm based on the principal component pursuit(PCP).The algorithm employs PCP to decompose the coefficient matrix of the MPLS model to obtain a low-rank structure of the normal data,and then rebuilds the MPLS model based on this low-rank structure.In this way,the influence of outliers in the input and output data is eliminated,and the robust model is built.Based on the robust MPLS model,monitoring indicators are developed to evaluate the quality-relevant subspace and the quality-irrelevant subspace.Numerical simulations and the Tennessee Eastman challenge problem are used to illustrate the effectiveness of the proposed method.(2)For the missing data in industrial processes,this dissertation introduces an iterative algorithm to estimate the missing data,and proposes a quality-related fault detection method based on the IA-RMPLS model.This method develops the robust MPLS by performing iterative alternation of establishing the IA-RMPLS model and estimating the missing data to model the robust MPLS.On this basis,the quality-irrelevant subspace is decomposed into a quality-irrelevant principal component subspace and a residual subspace based on the PCA method.Furthermore,the model is applied to quality-relevant fault detection,and monitoring indicators are constructed.The efficiency of the proposed method is demonstrated through its application in a numerical example and the TE process.(3)To overcome the heavier tail caused by cross and irrelevant items contained in the reconstruction-based contribution(RBC),this dissertation improves the RBC method and proposes a new concept called“contribution analysis”,and applies it to the robust MPLS monitoring model.Contribution analysis calculates the contribution of variables to the monitoring indicators_cT~2 and_rT~2,and forms a bounded convex hull to describe the contribution of variables to the fault.This method filters the fault variable set by eliminating the variables that do not contribute to the convex hull area,and finally obtains the main fault variables.Through the TE process,it is verified that the proposed method is more accurate than RBC recognition results. |
