Contribution au pronostic de la qualité d’un produit usiné par la proposition d’un modèle basé sur l’apprentissage machine

Abstract

This thesis is a contribution to the monitoring and prognostic of the quality of a machined workpiece based on a machine learning approach leveraging a variational autoencoder architecture. Still today, the manufacturing industry is prone to multiple pressures such as the globalization of markets, constant cost decrease requests or the ever-increasing quality-level requirements. Regarding the latter, this thesis’ objective is to demonstrate the possibility of developing an alternative measurement system capable of realizing the prognostic of the quality measurement value of a machined workpiece in real time. In this context, our first contribution was to propose a feature based on the concept of specific cutting energy. This feature is defined as the amount of energy required to remove 1 cm3 of raw material. It was demonstrated that this feature is highly correlated to a cutting tool wear’s level (r > 90%). We also shown that this feature has a significant contribution (14.7%) during the learning process in a supervised learning context. Another contribution of this thesis comes from our proposal of a model capable of making the prognostic of a workpiece’ quality measurement value. This model is based on a variational autoencoder type of neural network. The model is able to predict the quality values with a mean square error of 5.257 X 10‾⁴mm. The two dimensions latent space generated by the model visually distributes the data according to the quality level. Also based on this latent space and the Euclidean distance concept, we proposed a new metric capable of quickly quantifying the quality level. This metric is correlated with the observed quality values (r ≈ 94%) and with the predicted quality values (r ≈ 67%). Consequently, this latent space is a simple and visual tool that can be used to follow the evolution of the production process. This work was also accomplished in partnership with an industrial: APN Inc. (Quebec, Canada). Therefore, the results encompassed in this thesis are mostly based on a dataset acquired at the partner’s facility and representing more than 600 hours of regular production. This increases the potential of technological transfer to the industry and will also allows companies to better realize the monitoring and prognostic of the quality of their machining process. Furthermore, this opens the door to the possibility of implementing and using an alternative measurement system in the manufacturing industry.

Date	10 Jul 2021
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Antoine Tahan (Supervisor) & Marc Thomas (Co-supervisor)