Modèles d’apprentissage automatique pour un système manufacturier intelligent : Application au cas du pilotage d’un processus de séchage de bois

Abstract

Nowadays, Industry is strongly evolving and particularly the use of machine learning for the valorization of industrial data. This research thesis studies a drying process control problem in a wood boards production line using machine learning models. The moisture content (MC) is a crucial parameter to define the final value of the boards. Therefore, they must be dried to decrease their MC in order to meet customers’ requirements. The drying process in this research thesis uses a combination of two technologies: a conventional batch dryer, and a continuous high frequency (HF) dryer used as a precision kiln. The objective of this research is to use machine learning models to control the MC in the conventional dryer and at the entrance of the HF kiln in order to contribute to the implementation of an intelligent manufacturing system for the control of this process. Two sub-objectives have been set for this research thesis The first sub-objective is to use machine learning models to predict the mean MC in the conventional dryer to control the drying downtime. The prediction was made for every five minutes with ten hours lag. Several machine learning models were tested. A combination of a convolution layer with a bidirectional LSTM gave the best results with an R2 of 95.24% and an average absolute error of 3.61%. The second sub-objective deals with the prediction of the probability distribution of the MC at the entrance of the HF dryer for each package of the dried batch. This will help to determine the right packages to process in order to maximize the capacity of the HF dryer. The distribution estimation was based on the prediction of objective probabilities using multioutput predictive models. A multilayer perceptron, enhanced with upstream auto-encoders, gave the best results with a KL divergence of 0.53. These two predictions will enable the drying loop to be driven by controlling the conventional dryer's downtime and the distribution of packages’ MC at the HF kiln inlet.

Date	23 Feb 2023
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Mustapha Ouhimmou (Supervisor) & Loubna Benabbou (Co-supervisor)