Research On Nonlinear Quality Related Fault Diagnosis Method Based On Dynamic Latent Variable Model

A major industrial safety accident may seriously damage people's property,life and even the natural ecological environment.Therefore,fault diagnosis technology and application of large-scale industrial production processes are particularly important.The data-driven fault diagnosis method is a hot research topic in recent year which can overcome the limitation of the physical model and realize the fault diagnosis more accurately.Among them,the multivariate statistical analysis method is very suitable for large-scale industrial production processes,which need simple modeling process,small amount of calculation,and use normal sample for training.Based on the key performance indicators that people are most concerned about,this thesis deeply studies the fault diagnosis methods of multivariate statistical analysis for the dynamic and nonlinear characteristics contained in the actual production process.This thesis proposes dynamic system quality variable prediction,nonlinear system quality-related fault detection and fault diagnosis method:1.To deal with the ubiquitous dynamic characteristics in actual industrial systems,a dynamic concurrent partial least squares model is proposed to reduce the complexity and calculation amount of dynamic system data modeling,which achieve accurate prediction of quality variables in dynamic system.Combining the inner model and outer model of the dynamic concurrent partial least squares model,the objective function is constructed and solved by iteration.After that,the Lagrangian multiplier method is used to prove that the objective function has a global optimal solution.2.In order to solve the problem of nonlinear system quality-related fault detection,a kernel least squares model is proposed which decompose the high-dimensional space into two orthogonal subspaces to realize the nonlinear system fault detection.Based on this,a partial derivate contribution plot is proposed which expresses the fault intensity by calculating the kernel gradient instead of the contribution value.3.To reduce the smearing effect which is common in the fault diagnosis method of multivariate statistical analysis,a reconstructed partial derivate contribution plot is proposed which can suppress smearing effect.This method uses the idea of data reconstruction to preprocess the fault samples before fault diagnosis which can eliminate the smearing effect of fault variables to normal variables.In addition,a more reasonable fault diagnosis control limit is designed in reconstructed partial derivate contribution plot based on the distribution of the contribution values of the process variables of the training samples,which significantly optimizes the diagnosis effect.Numerical simulations and Tennessee-Eastman process are used to verify the effectiveness of the proposed methods.Finally,the main research content of the thesis is summarized,and the direction and topics of continuing research are discussed.