Ph. D. Project
Foreseeing sawmills' production through 3D cloud points analysis and machine learning algorithms
2020/10/01 - 2023/09/30
Other supervisor(s):
Jonathan Gaudreault (
In many regards, the forest-wood supply chain is not unlike that of the companies of other sectors (products
or services, suppliers to customers, sales). There are, however, key differences in the first few steps of the
sawing process rendering the planning phase particularly difficult. The raw material, i.e., the harvested trees,
is inherently heterogeneous. Sawing a given section of a felled tree causes the generation of several products
at the same time. Production planning, in this context, is difficult and is mostly, to this date, based on on-site
production systems. The storage costs, on the one hand, and the randomness of sales, on the other hand, also
increase the difficulty of forecasting and managing the costs and the sales.

The goal of the proposed project is to harness the power of novel prediction and simulation technologies
based on 3D imaging to optimize the allocation of wood to sawmills. The approach will, on the long-term,
help to put into perspective the current stock-based manufacturing approach widely used in the wood
industry by improving the forecasting capabilities of the companies. To meet this objective, French and
Canadian laboratories (CRAN, FORAC) will join forces each bringing a complementary expertise. CRAN's
research efforts and recent works (Almecija, 2013) (Jover et al., 2013) made the development of new wood
identification and characterization approaches possible. Those can be, for instance, advantageously coupled
with decision support tools such as virtual sawing improving their efficiency and precision. FORAC recently
explored the use of machine learning to predict the products resulting from the sawing of a log at a given
sawmill (Morin et al., 2019). Those predictions were based on the characteristics of a log, e.g., its length and
its curvature, which is a summary of the information contained in its 3D scan, i.e., a point-cloud. In recent
works, the two laboratories explored methods aimed at exploiting the entire information contained in the 3D
scans for forecasting purposes, an information which is hard to integrate in learning algorithms made for fixed
size input such as k-nearest neighbors, random forest, and decision tree. Preliminary results (Selma et al.,
2017) obtained using iterative closest point methods from 3D imaging, enabled the team to show how the
similarity between two input logs described by their complete 3D scans can be established. Given the 3D scan
of a log for which the output is known and the 3D scan of a log for which the output is unknown, such a
similarity measure renders it possible to know whether or not the logs resemble each other and as a
consequence whether or not output of the second is likely to be close to the output of the first. Such a
measure, based on the complete scan, is likely to be more accurate than a measure based on general
characteristics since it uses the entire information contained in the 3D scans.

In this project, we propose point-cloud normalization and/or filtering to improve the accuracy of the 3D scans,
our input data for similarity evaluation, which comes from low-cost scanners. This data, once processed, will
be used by machine learning algorithms to predict the output of the sawing of the logs at sawmills. The
proposed learning algorithm will combine 3D matching distance calculation with classical machine learning
approaches. The distance calculation procedure is, nonetheless, expensive. We plan to mitigate that cost by
using a first classification step based on the general characteristics of the logs and a second, more accurate,
classification step using the developed similarity measure based on the complete normalized/filtered 3D
scans. Learning algorithms such as k-nearest neighbors, neural networks, decision trees and random forest
will be evaluated for the first and second steps.

The scientific challenges of that project include the variation in the size of the 3D scans which complicate a
direct use of machine learning on that input, i.e., the number of points is not equal across scans, the
improvement of our current similarity models by point-cloud filtering/normalization approaches, the
development of the hierarchical classification procedure based on coarse/precise input data as described
above, and the parameterization of the different algorithms such that the best possible
precision/performance compromise is achieved.
References :
Morin M, Gaudreault J, Brotherton E, Paradis F, Rolland A et. al. Machine learning-based models of sawmills for better wood allocation planning.
International Journal of Production Economics
2019 pp: 107508
Cyrine Selma, Hind El Haouzi, Philippe Thomas, Jonathan Gaudreault, Michael Morin. An iterative closest point method for measuring the level of
similarity of 3D log scans in wood industry, SOHOMA 2017, Oct 2017, Nantes, France. 2017
3D cloud points analysis, Machine Learning, Classification, Prediction/Simulation
3 years contract,

We're looking for a candidate holding a master degree in computer science/industrial engineering with strong
competence on computer science and in particular on maching learning algorithms
Modeling and Control of Industrial Systems
ANR Phd grants