Automated manufacturability analysis and machining process selection using deep generative model and Siamese neural networks

Abstract

Industry 4.0 calls for highly autonomous manufacturing process planning. Significant effort has been devoted over the years to generative Computer Aided Process Planning (CAPP), which aims to generate process plans for new designs without human intervention. This goal has not been realized to date due to several reasons, such as poor scalability and the difficulty in modeling manufacturing process capability, which encapsulates the part shape, quality, and material property transformation capabilities of the process. In our prior work, the shape transformation capabilities of lathe-based machining operations were modeled as latent probability distributions using a data-driven deep learning-based generative machine learning approach, from which visualizations of realistic machinable features could be sampled to assist manual process selection. In this paper, a Siamese Neural Network (SNN) is integrated with Autoencoder-based deep generative models of machining operations to enable automated comparison of the query part shapes with sampled outputs. This enables automated manufacturability analysis and machining process selection necessary for generative CAPP. The paper also demonstrates that the proposed Autoencoder and Siamese Neural Network (AE-SNN) achieves a class-average process selection accuracy of 89 %, and a manufacturability analysis accuracy of 100 %, which outperforms a discriminative model trained on the same dataset.