Ph. D. Project
Title:
Machine learning based on data from the panel manufacturing process for prediction and optimization of the mix-
products and energy flow
Dates:
2021/10/01 - 2024/09/30
Supervisor(s):
Description:
This thesis topic is part of the scientific project S&O-2I of the Eco-technical Systems Engineering Department of
CRAN. This work will be carried out in collaboration with the EGGER company which is one of the European
leaders in the wood panel market.
The complexity of the planning and control system of the particleboard production process is due to the fact
that particleboard is a basic product. Therefore, price is a major determinant of the manufacturer's
competitive position. However, the panels are subject to high normative requirements depending on the areas
of use, but also to specific customer requirements. The quality of particleboard depends on many variables, in
particular the composition of its raw materials, i.e. the wood species used and the types of wood raw material
processed (logs, debarked slabs, sawdust, etc.). We will call this composition the product mix.
In this thesis, we will focus on the influence of this product mix on the cost price, energy consumption and
other performances that remain to be identified in a sustainable development logic advocated by the
company. This optimization of the product mix depends on the availability of a given wood raw material, on
specific quality requirements, on the processing capacities of the plant...
The complexity and dynamic nature of this problem, as well as the dominance of the empirical experience of
operators and managers makes it an ideal candidate for the implementation of a decision support system. The
latter will have to help the manager find a solution that maximizes the performance of the product mix.
The industrial problem introduced above is of interest to the laboratory of the Centre de Recherche en
Automatique de Nancy (CRAN) and more precisely to the department of Engineering of Complex Eco-Technical
Systems (ISET). This department is interested in the modeling of evaluation and decision-making processes
necessary to control and pilot complex systems.
The thesis will aim to propose models for prediction and optimization of the influence of the product mix in
the planning process with a multi-dimensional performance research, exploiting the data collected on the
manufacturing process.
The project can rely on the various works carried out at CRAN and in particular within the ISET department for
several years on the exploitation of industrial data to build models using machine learning techniques in
general and neural networks in particular. These activities have focused both on the development of new
algorithms to improve the accuracy of learned models and on the application of these techniques to industrial
problems. It will also be able to draw on the work of
Based on the experiences and knowledge of the laboratory and the results of a first preliminary study within
the framework of the PFE -ENSTIB-EGGER 2020, we have identified several locks to be lifted :
- the use of heterogeneous data. This heterogeneity of data is a problem in its own right when using machine
learning techniques because few of them are adapted to the joint use of discrete and continuous data.
- pollution of data by noise and outliers. It is necessary to implement techniques and tools capable of ensuring
the good accuracy of the model in spite of these perturbations. In the context of neural networks, for example,
this requires the use of robust learning algorithms in the presence of outliers, as well as the determination of
the optimal structure of the network to avoid overfitting problems,
- The complexity of the flows to be considered, heterogeneous raw material flows with several loops at the
drying and grinding level, but also the energy flow (used for equipment consumption, waste management and
sales to customers) also leads us to focus on simulation models and the coupling of the latter with learning
models in a digital twin logic.
- deployment on other sites. All the models we will produce will be developed on the basis of data from a
single site. The idea is to be able to deploy the tools thus built on other EGGER group sites.
CRAN. This work will be carried out in collaboration with the EGGER company which is one of the European
leaders in the wood panel market.
The complexity of the planning and control system of the particleboard production process is due to the fact
that particleboard is a basic product. Therefore, price is a major determinant of the manufacturer's
competitive position. However, the panels are subject to high normative requirements depending on the areas
of use, but also to specific customer requirements. The quality of particleboard depends on many variables, in
particular the composition of its raw materials, i.e. the wood species used and the types of wood raw material
processed (logs, debarked slabs, sawdust, etc.). We will call this composition the product mix.
In this thesis, we will focus on the influence of this product mix on the cost price, energy consumption and
other performances that remain to be identified in a sustainable development logic advocated by the
company. This optimization of the product mix depends on the availability of a given wood raw material, on
specific quality requirements, on the processing capacities of the plant...
The complexity and dynamic nature of this problem, as well as the dominance of the empirical experience of
operators and managers makes it an ideal candidate for the implementation of a decision support system. The
latter will have to help the manager find a solution that maximizes the performance of the product mix.
The industrial problem introduced above is of interest to the laboratory of the Centre de Recherche en
Automatique de Nancy (CRAN) and more precisely to the department of Engineering of Complex Eco-Technical
Systems (ISET). This department is interested in the modeling of evaluation and decision-making processes
necessary to control and pilot complex systems.
The thesis will aim to propose models for prediction and optimization of the influence of the product mix in
the planning process with a multi-dimensional performance research, exploiting the data collected on the
manufacturing process.
The project can rely on the various works carried out at CRAN and in particular within the ISET department for
several years on the exploitation of industrial data to build models using machine learning techniques in
general and neural networks in particular. These activities have focused both on the development of new
algorithms to improve the accuracy of learned models and on the application of these techniques to industrial
problems. It will also be able to draw on the work of
Based on the experiences and knowledge of the laboratory and the results of a first preliminary study within
the framework of the PFE -ENSTIB-EGGER 2020, we have identified several locks to be lifted :
- the use of heterogeneous data. This heterogeneity of data is a problem in its own right when using machine
learning techniques because few of them are adapted to the joint use of discrete and continuous data.
- pollution of data by noise and outliers. It is necessary to implement techniques and tools capable of ensuring
the good accuracy of the model in spite of these perturbations. In the context of neural networks, for example,
this requires the use of robust learning algorithms in the presence of outliers, as well as the determination of
the optimal structure of the network to avoid overfitting problems,
- The complexity of the flows to be considered, heterogeneous raw material flows with several loops at the
drying and grinding level, but also the energy flow (used for equipment consumption, waste management and
sales to customers) also leads us to focus on simulation models and the coupling of the latter with learning
models in a digital twin logic.
- deployment on other sites. All the models we will produce will be developed on the basis of data from a
single site. The idea is to be able to deploy the tools thus built on other EGGER group sites.
Keywords:
intelligent manufacturing control, energy and mix-products predection an optimization, Digital TWIN
Department(s):
| Modeling and Control of Industrial Systems |
Funds:
CIFRE Funding
Publications:
