Root Cause Diagnosis And Propagation Path Identification Of Quality-related Faults For Complex Industrial Processes

Associated with the increasing demands on the improvements of quality,efficiency and benefits from modern industrial processes,this thesis is mainly concentrated on root cause diagnosis and propagation path identification of quality-related faults in industrial processes,based on traditional multivariate statistics,probabilistic analysis,machine learning,information fusion,graph theory and pattern recognition techniques.Meanwhile,the problems of long process flow,multiple operation conditions,dynamic,nonlinearity,and difficult location,complex propagation path and evolution characteristics of quality-related faults for complex industrial processes have been fully considered.This thesis intends to provide theoretical basis and technical support for reducing quality-related faults,ensuring the product quality and improving the economic benefits.To be specific,the main innovations are as follows:1)To solve the problems such as complex modeling,wide-ranged and difficult location of quality-related faults,and inefficient causal topology construction for complex industrial processes,a root cause diagnosis scheme is proposed based on combination of candidate dataset selection and causality analysis.Under that scheme,for dynamic process,a mutual information canonical variable analysis model is built first for quality-related fault detection.Then,the generalized reconstruction based contribution is investigated to select the candidate dataset of faulty variables.Finally,a transfer entropy-based causality analysis method is put forward for root cause diagnosis of quality-related faults.For multimode process,a robust Gaussian mixture model is built for quality-related fault detection.Then,a Bayesian inference-based robust Gaussian mixture contribution index is designed to analyze the potential root-cause variables.Finally,a combination of transfer entropy and direct transfer entropy-based causality analysis method is proposed for root cause diagnosis of quality-related faults.2)Due to the fact that traditional data-based causal topology construction methods may have more redundant connections,and knowledge-based methods may lead to losses of large amounts of non-intuitive and important information,based on process knowledge and data,a propagation path identification strategy is presented.Under that strategy,for nonlinear dynamic process,a nonlinear dynamic latent variable model is built first for quality-related fault detection.Then,the relative reconstruction based contribution approach is investigated to analyze the potential root-cause variables.Finally,a partitioned Bayesian network methodology is proposed for propagation path identification of quality-related faults.For dynamic process,on the basis of local and global quality-related fault detection results based on data-driven gap metric and Bayesian inference,a neural network architecture-based Granger causality analysis method is developed for propagation path identification of quality-related faults.3)Since the global modeling methods may ignore many process details,and the varying,unknown and multimode characteristics of quality-related multiple faults,a top-down hierarchical detection and root cause diagnosis framework is designed for quality-related multiple faults.Under that framework,for dynamic process,on the basis of local and global quality-related fault detection results based on correlation-based canonical variable analysis and Bayesian inference,a tensor subspace analysis-based discriminant analysis approach is proposed for locating the root causes.For nonlinear dynamic process,on the basis of local and global quality-related fault detection results based on adaptive kernel canonical variable analysis and Bayesian inference,a robust sparse exponential discriminant analysis algorithm is proposed for root cause diagnosis of unknown and multimode quality-related multiple faults.The above research results provide new ideas and solutions to address the key issues on root cause diagnosis and propagation path identification of quality-related faults in complex industrial processes.Moreover,the proposed methods are verified in the hot strip mill process and Tennessee Eastman process,and are compared with several monitoring and diagnosis methods to demonstrate the advantages of these new approaches.