Please use this identifier to cite or link to this item:
|Title:||Diagnosis of faults in plant operations and scheduling under uncertain production times|
|Abstract:||Technical advances since 1970s in the areas of plant measurements and data archiving have enabled us to measure large number of process variables and record in the industrial plant historians. Early stages of plant data analysis via charting have been replaced by methods which enable us to extract more useful information from plant data, thereby providing valuable insight into plant operations. Monitoring of plant operations needs enables us to assess whether production targets are met according to the schedule. Both of these activities, monitoring and scheduling, are subject to uncertainties. Uncertainty in the monitoring arises from the possibility that some measurements may not be accurate or that the process equipment has developed a fault. Uncertainty in scheduling is caused by variations in the length of time required to complete individual tasks in the multi-step production system. This Thesis deals with both of the above aspect. The first part introduces novel process monitoring and fault diagnosis methods which have been developed to extract useful information from highly correlated process variables. The second part focuses on modeling and optimizing manufacturing processes impacted by many uncertainties, which necessitates statistical techniques. Multivariate statistical process monitoring (MSPM) techniques developed in this research extract useful information from a large number of highly correlated process variables and historical data sets. In order to capture the non-Gaussian features and relationships between input and output variables, a new quality relevant non-Gaussian latent subspace projection method is proposed by adopting the high-order statistics of mutual information for searching the latent directions within input and output spaces, respectively. Moreover, to monitor quality related operational performance of nonlinear batch processes, a novel multiway kernel based quality relevant non-Gaussian latent subspace projection method is developed. Furthermore, a new probabilistic graphical model based network process monitoring has been developed for the identification of the root-cause variable. Modeling of uncertainties in production times and scheduling under such uncertainties are subject of the research in the second part of this Thesis. The Bayesian network models are proposed for accurate estimation of the production loads and the total production times in manufacturing processes. The proposed models are applied to schedules for steelmaking continuous casting production. Continuous casting scheduling is a difficult optimization problem due to a large number of binary variables that are needed to represent exactly process characteristics in the optimization model. To solve the scheduling problems, a new two-level algorithm and parallel simulated annealing methods are developed. Moreover, the proposed algorithm is extended to a multi-objective evolutionary algorithm in order to optimize simultaneously multiple objectives. Real life steel production process data have been used to examine the effectiveness of all the above proposed algorithms.|
|Appears in Collections:||Open Access Dissertations and Theses|
Files in This Item:
|Mori_Junichi_finalsubmission2016August_PhD.pdf||14.86 MB||Adobe PDF||View/Open|
Items in MacSphere are protected by copyright, with all rights reserved, unless otherwise indicated.