Search

Features of jumping loads are essentially high-dimensional random variables but have been simplistically modeled owing to the lack of proper mathematical tools. Generative adversarial networks (GANs) in conjunction with deep learning technology are adopted herein for modeling the jumping loads. Conditional GANs (CGANs) combined with Wasserstein GANs (WGANs) with gradient penalty (WGANs-GP) are adopted in the impulse modeling, where a multi-layer perceptron and a convolutional neural network are employed for the discriminator and generator, respectively. As for the impulse amplitude sequence and interval sequence modeling, similar CGANs combined with WGANs-GP are adopted, where recurrent neural networks are employed for both the generator and discriminator. A large amount of measured individual jumping loads are utilized in training GANs to ensure the generated samples can simulate the real ones well. After training, the individual jumping loads are simulated by connecting the generated impulse samples, based on the generated impulse amplitude sequence samples and interval sequence samples. The simulated jumping loads can be used to assess the vibration performance of assembly structures, such as grandstands, concert halls, and gym floors. Moreover, the established GANs can be extended to the modeling of other stochastic dynamic excitations. [ABSTRACT FROM AUTHOR]

Strong mainshocks are typically accompanied by numerous aftershocks, and the investigation of the structural failure mechanisms under the mainshock-aftershock sequence becomes particularly crucial. However, the number of recorded mainshock-aftershock sequences is limited. Therefore, the purpose of this article is to provide a reasonable method for directly generating the aftershock time histories from mainshock time histories. Using convolutional network as the basic network layer and conditional generative adversarial network as the structure, two models, one-dimensional convolution (1D-C-DCGAN) and two-dimensional convolution (2D-C-DCGAN) are established respectively by utilizing the deep convolutional generative adversarial network to learn the relationship between the mainshock-aftershock time histories. Then, they are trained with 972 pairs of the selected mainshock-aftershock time histories, and prediction results are discussed in comparison. The results show that the two models are proficient in generating AS acceleration time histories that are closely related to the sample trend, in which the 2D-C-DCGAN model performing better in overall waveform prediction, but with local spikes. In the comparison of intensity measures and response spectra, by examining coefficients such as R2, RMSE, MAPE, the two models outperformed the mainstream model (ASK14) on the dataset, and the 2D-C-DCGAN model is more accurate than the 1D-C-DCGAN model. The distributions of intensity measure predicted by 2D-C-DCGAN model are closer to the measured intensity measures, and its predicted response spectra are smoother and better matched to the measured response spectra. This advantage can be attributed to the effectiveness of convolution operations on two-dimensional data, allowing the convolutional capabilities to be fully utilized. [ABSTRACT FROM AUTHOR]