Researches On Fault Diagnosis And Quality Prediction For Steel Rolling Process Based On Multivariate Statistical Analysis

In the steel rolling process, the effective fault diagnosis and quality prediction are the keys to ensure production safety, enhance product quality and increase economy benefit. However, it is difficult to construct the accurate mathematical models for the complex steel rolling process. These limit the further research and application of fault diagnosis and quality prediction methods based on the mathematical model. Considering the characteristics of the steel rolling process in the followings:(1) The structure is very complex;(2) a large amount of process data are sampled, collected and obtained;(3) process data is nonlinear, dynamic and does not obey Gaussian distribution, In this thesis, fault diagnosis and quality prediction methods are researched in the steel rolling process based on multivariate statistical analysis methodologies in depth. Multivariate statistical analysis is a data-driven method, which can extract process information feature, determine operating status of the process, diagnose the abnormal condition and fault, and then predict product quality through analysis and interpretation of the collected measurements.In terms of application, this thesis considers comprehensively several key problems need to be solved when multivariate statistical methods are applied in the steel rolling process and develops a series of fault diagnosis and quality prediction methods for the steel rolling process. The main results and contributions of this thesis are stated as follows:(1) A relative-transformation partial least square method based on Mahalanobis Distance (MRTPLS) for condition monitoring of steel rolling process is proposed. Firstly, transferring the original data space into the relative space by computing Mahalanobis Distance between sample data. Secondly, using PLS approach in the relative space to extract the representative latent variables, building the statistical monitoring model and then monitor condition of steel rolling process. Both theory analysis and simulation example demonstrate that MRTPLS can remove directly effect of dimension, extract effectively the latent variables with better variation degree and representativeness, and improve the accuracy and real-time of condition monitoring. (2) In view of the characteristics of dynamic, nonlinear and non-Gaussian distribution in steel rolling process, a combination method of improved dynamic kernel principle component analysis and independent component analysis for condition monitoring and fault isolation is proposed. Firstly, constructing augmented data matrix and separates into some sub-augmented matrices. Secondly, applying kernel principle component analysis to each sub-matrix for extracting the nonlinear cross-correlation feature, respectively. Thirdly, using all nonlinear principal components to construct a new augmented data matrix. At last, building independent component analysis statistical model so that the condition of steel rolling process is monitored efficiently. Moreover, developing a new contribution plot method based on nonlinear independent component which maintains the advantages of the simplicity of contribution plots method as the same time, and meanwhile improves the accuracy of fault isolation.(3) A kernel Fisher discriminant analysis(KFDA) method based on optimization strategy for condition monitoring and fault identification is proposed. The proporsed method uses improved biogeography-based optimization approach to optimize kernel parameters and select feature data sample simultaneously. And then using the optimal kernel parameters and feature sample to build KFDA statistical modeling, monitor process condition and identify fault type, so that the computational efficiency of kernel matrix and performance of on-line condition monitoring and fault identification.(4) A fault diagnosis and quality prediction method based on nonlinear feature extraction and regression is proposed. This method firstly extracts the nonlinear feature of process data by kernel PLS, establishes internal regression model of KPLS, and then obtains optimal discriminant feature vector which satisfies maximal separation degree for condition monitoring. If the process is under normal condition, the regression model of KPLS is further used for quality prediction. Otherwise, the similar degree in the fault feature direction is used to identify fault types.In the course of the study, a series of of fault diagnosis and quality prediction proposed in the thesis are studied by a large number of simulation experiments making full use of the field actual data. The validity and feasibility of these methods are verified through the experimental results.