Nowadays, aluminium products, especially extruded profiles, are widely used in a variety of fields, such as transportation and construction. The great popularity of these profiles can be explained by their relatively low manufacturing cost as well as a strong versatility in the shape and size of the produced parts. This allows the production of specific profiles for each customer. Unfortunately, this versatility creates an additional difficulty: every request for new designs must be analyzed in detail to determine their feasibility. Several aspects have to be considered such as the alloy, shape and dimensions of the profile. Obviously, there is no fixed formula to determine whether a profile is feasible or not, this study often relies heavily on the expert’s knowledge accumulated from many years of practice.
Thus, this thesis presents an approach based on Artificial Intelligence (AI) and more particularly on Machine Learning (ML) in order to predict the feasibility of the extrusion of these profiles. The latter being based on technical drawings of the profiles sections as well as on 11 geometric data in scalar format, three Convolutional Neural Network (CNN) models were proposed and implemented. Four target outputs of these networks were determined in order to represent the manufacturability score, namely the productivity, the success rate, the rectification rate and the rejection rate of the produced parts. The models were trained on datasets containing more than 1600 entries, collected from the historical production data of Constellium Valais. Finally, an interface integrating the best model was implemented in order to offer a simple and fast predictive tool. However, the models trained with the available data allowed to obtain a reliable prediction for only one of the four chosen targets: the productivity.