| Modern industrial system has the characteristics of large scale,multi-complex coupling of control variables,and multi-modes,which makes it extremely difficult to carry out fault detection on the system based on knowledge and basis modeling.However,the data-driven methods represented by machine learning,multivariate statistical analysis and signal processing have the advantage that they do not need to conduct in-depth research on the system process.Principal component analysis(PCA)plays an important role in large-scale industrial fault detection in recent years because of its ability to reduce the dimension of complex industrial process data,retain the data that can best reflect the characteristics of the system and simplify the analysis of complex data.However,the principal component analysis method requires the system process data to be single-mode,and for modern industry,multi-mode process switching is normal.Therefore,this thesis studies the multi-mode fault detection method based on the principal component analysis,and the main contents include:1.The characteristics of multi-mode data are analyzed.For the multi-modal data preprocessing by z-score method cannot change the multi-modal characteristics of the data,A Double Neighbor Weighted Preliminaries-summation Standardization(DNWPS)method is proposed.After preprocessing multimodal data with this method,The data is changed from multi-mode to single-mode approximately Gaussian distribution,and the validity of this method is verified by numerical experiments.2.The characteristics of data preprocessed by DNWPS method are analyzed.In view of the high dependence of the statistic based fault detection method commonly used in principal component analysis on the Gaussian property of data,this thesis proposes a fault detection method based on WOA-SVDD;In order to solve the problem of slow training speed of WOA-SVDD method,a clustering method based on data density was proposed to reduce training samples and then improve the training speed of WOA-SVDD.The validity of the sample reduction method is verified by the TE platform dataset experiment.3.A complete multi-mode fault detection method flow based on principal component analysis is proposed,and multi-mode step fault and slow drift fault are simulated numerically.Four kinds of typical multimodal faults are simulated by TE platform:step fault,slow drift fault,random transformation fault and unknown fault.Through the above two simulation methods,it is verified that when dealing with multi-mode problems,the global weighted nearest neighbor summation method is firstly used to preprocess the data,then the principal component analysis method is used to reduce the dimension,and finally the WOA-SVDD method after sample reduction is used to detect the fault samples.It is better than the z-score preprocessing principal component analysis method and the local nearest neighbor standardized principal component analysis method proposed by other scholars. |
