Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
BallastGAN: Random generation of ballast particle contour based on generative adversarial networks
https://orcid.org/0000-0003-1527-6619
Abstract Railway ballast particles exhibit unique variations in shape, and this diversity significantly impacts the mechanical performance of ballast beds. An intuitive approach involves defining corresponding morphological parameters through morphological analysis of particle contours to randomly generate ballast particle contours. This involves deriving the probability distribution of these parameters based on mathematical statistics of ballast particle samples. However, due to the empirical nature and underlying assumptions of these morphological parameters, the accuracy of the generated particle contours is often compromised. To generate highly realistic ballast particle images, we propose a novel model named BallastGAN for random ballast particle contour generation. This model employs fractal dimension regularization and optimizes the loss function of the discriminator within the framework of StyleGAN3. This optimization effectively controls the roughness characteristics of generated ballast particle images. Mutual transformation and style blending between different ballast particle contours are achieved by decoupling the latent space. The ballast particle image dataset (BPID) is curated to train the model. Experimental results demonstrate that the proposed model's randomly generated ballast particle contours exhibit a remarkably high degree of consistency with real contours. This indicates that the generated contours can be directly applied to numerical simulations in mechanics.
Highlights Random generation method of ballast particle image based on deep learning. Fractal dimension normalization is applied in the image generation model of ballast particles. Engineering application of generative adversarial networks.
BallastGAN: Random generation of ballast particle contour based on generative adversarial networks
https://orcid.org/0000-0003-1527-6619
Abstract Railway ballast particles exhibit unique variations in shape, and this diversity significantly impacts the mechanical performance of ballast beds. An intuitive approach involves defining corresponding morphological parameters through morphological analysis of particle contours to randomly generate ballast particle contours. This involves deriving the probability distribution of these parameters based on mathematical statistics of ballast particle samples. However, due to the empirical nature and underlying assumptions of these morphological parameters, the accuracy of the generated particle contours is often compromised. To generate highly realistic ballast particle images, we propose a novel model named BallastGAN for random ballast particle contour generation. This model employs fractal dimension regularization and optimizes the loss function of the discriminator within the framework of StyleGAN3. This optimization effectively controls the roughness characteristics of generated ballast particle images. Mutual transformation and style blending between different ballast particle contours are achieved by decoupling the latent space. The ballast particle image dataset (BPID) is curated to train the model. Experimental results demonstrate that the proposed model's randomly generated ballast particle contours exhibit a remarkably high degree of consistency with real contours. This indicates that the generated contours can be directly applied to numerical simulations in mechanics.
Highlights Random generation method of ballast particle image based on deep learning. Fractal dimension normalization is applied in the image generation model of ballast particles. Engineering application of generative adversarial networks.
BallastGAN: Random generation of ballast particle contour based on generative adversarial networks
https://orcid.org/0000-0003-1527-6619
Abstract Railway ballast particles exhibit unique variations in shape, and this diversity significantly impacts the mechanical performance of ballast beds. An intuitive approach involves defining corresponding morphological parameters through morphological analysis of particle contours to randomly generate ballast particle contours. This involves deriving the probability distribution of these parameters based on mathematical statistics of ballast particle samples. However, due to the empirical nature and underlying assumptions of these morphological parameters, the accuracy of the generated particle contours is often compromised. To generate highly realistic ballast particle images, we propose a novel model named BallastGAN for random ballast particle contour generation. This model employs fractal dimension regularization and optimizes the loss function of the discriminator within the framework of StyleGAN3. This optimization effectively controls the roughness characteristics of generated ballast particle images. Mutual transformation and style blending between different ballast particle contours are achieved by decoupling the latent space. The ballast particle image dataset (BPID) is curated to train the model. Experimental results demonstrate that the proposed model's randomly generated ballast particle contours exhibit a remarkably high degree of consistency with real contours. This indicates that the generated contours can be directly applied to numerical simulations in mechanics.
Highlights Random generation method of ballast particle image based on deep learning. Fractal dimension normalization is applied in the image generation model of ballast particles. Engineering application of generative adversarial networks.
BallastGAN: Random generation of ballast particle contour based on generative adversarial networks
Wang, Yang (Autor:in) / Xiao, Hong (Autor:in) / Chi, Yihao (Autor:in) / Zhang, Zhihai (Autor:in) / Qian, Zhongxia (Autor:in)
07.12.2023
Aufsatz (Zeitschrift)
Elektronische Ressource
Englisch
Structural Plan Schema Generation Through Generative Adversarial Networks
| Springer Verlag | 2024
Structural Plan Schema Generation Through Generative Adversarial Networks
| Springer Verlag | 2024
Controllable Generation of Porous Microstructures Through Generative Adversarial Networks
| Springer Verlag | 2025