Skip to Main content Skip to Navigation
Theses

Méthodes statistiques et variationnelles de modélisation préalable au contrôle de procédés industriels

Abstract : In this thesis, we aim to propose a methodology for analyzing industrial processes based on a large quantity of process data with statistical and variational methods. The objective is to identify the key factors which ensure the good functioning of industrial processes. It is a preliminary step for the automatic generation of process control law from the collected data. In the first part of this thesis, we presented the statistical analysis of a process of Saint-Gobain. At first, we applied some classical statistical tools (Principal component analysis, clustering and so on) to the process data, and linked the obtained information with the functioning of the process. Then, we analyzed a product quality measure (called target) with the collected process parameters. The target is weakly correlated with the parameters, so the hypothesis of linear model is rejected. A restricted list of parameters which contribute to the explaination of the target was identified by the statistical methods and validated by our industrial interlocutors. Afterwards, we tested a non-linear model method: the generalized additive model (GAM). The introduction of the non-linear terms improved the performance of our model, but it remained insufficient for the future applications. Following the intuition of the process engineers and operators, we focused on a noisy signal, tracked regularly in the plants, characterizing the good functioning of the process, and restoring the missing information of this signal may improve the model. In the second part of this thesis, we developed a total variation restoration method with an automatic choice of hyper-parameter. Furthermore, our proposition of hyper-parameter has a similar performance as the existing methods, and our estimation method of both hyper-parameter and restoration is well fitted for the real time processing of a large quantity of data. Based on the proposed method, we have developed some applications of pattern restoration and discontinuities detection for several industrial processes.
Document type :
Theses
Complete list of metadata

https://hal.univ-lorraine.fr/tel-03274049
Contributor : Thèses Ul <>
Submitted on : Tuesday, June 29, 2021 - 4:50:03 PM
Last modification on : Friday, July 2, 2021 - 6:07:39 PM

File

DDOC_T_2021_0086_LIU.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : tel-03274049, version 1

Collections

Citation

Zhanhao Liu. Méthodes statistiques et variationnelles de modélisation préalable au contrôle de procédés industriels. Mathématiques [math]. Université de Lorraine, 2021. Français. ⟨NNT : 2021LORR0086⟩. ⟨tel-03274049⟩

Share

Metrics

Record views

69

Files downloads

24