Proceedings of the
9th International Conference of Asian Society for Precision Engineering and Nanotechnology (ASPEN2022)
15 – 18 November 2022, Singapore
doi:10.3850/978-981-18-6021-8_OR-08-0053
Predicting 2D Data: Using conditional Generative Adversarial Networks in Incremental Sheet Forming
1Department of Mechanical Engineering, National University of Singapore, 9 Engineering Drive, Singapore, 117576, Singapore
ABSTRACT
Incremental Sheet Forming (ISF) is a flexible sheet forming process which utilizes a moving forming tool to incrementally deform a metal sheet into the desired geometry. The elastic recovery of the material after deformation causes a spring-back effect to occur during the process, which contributes to significant geometrical error in the final shape. In this paper, we present the use of a conditional Generative Adversarial Network (cGAN) trained using empirical data to predict such errors. Using only the 2D depth map of the target geometry, the network generated 2D predictions of the final geometry. This differs from other prediction models in that it doesn't require handcrafting of parameters describing the local geometry, making it more compatible with complex free-form geometries. Experimental investigations reveal that the network was able to generate predictions of untrained geometries with 83.9%-99.6% accuracy. Subsequently, we demonstrate a case study for the use of the prediction to improve geometric accuracy. This paper would be useful to researchers considering empirical modelling of 2D data.
Keywords: generative adversarial networks, machine learning, incremental sheet forming
1Department of Mechanical Engineering, National University of Singapore, 9 Engineering Drive, Singapore, 117576, Singapore
ABSTRACT
Incremental Sheet Forming (ISF) is a flexible sheet forming process which utilizes a moving forming tool to incrementally deform a metal sheet into the desired geometry. The elastic recovery of the material after deformation causes a spring-back effect to occur during the process, which contributes to significant geometrical error in the final shape. In this paper, we present the use of a conditional Generative Adversarial Network (cGAN) trained using empirical data to predict such errors. Using only the 2D depth map of the target geometry, the network generated 2D predictions of the final geometry. This differs from other prediction models in that it doesn't require handcrafting of parameters describing the local geometry, making it more compatible with complex free-form geometries. Experimental investigations reveal that the network was able to generate predictions of untrained geometries with 83.9%-99.6% accuracy. Subsequently, we demonstrate a case study for the use of the prediction to improve geometric accuracy. This paper would be useful to researchers considering empirical modelling of 2D data.
Keywords: generative adversarial networks, machine learning, incremental sheet forming
