Near-fault ground motion synthesis based on conditional generation adversarial network

Near-fault (NF) ground motions usually have high-amplitude and long-period velocity pulses that might cause excessive responses in flexible structures. However, the number of recorded NF ground motions is very limited and hinders related research in earthquake engineering. In this paper, we develop a conditional generative adversarial network (CGAN) model, namely Ep2NgmGAN, to generate NF ground motions under given engineering parameters. Different from the traditional CGAN model, it inputs the label by introducing a label embedding module. In addition, a knowledge-enhanced module is adopted to enable the model to capture prior knowledge about NF ground motions. Using the strategy suggested in this study, the Ep2NgmGAN is trained and tested on the dataset constructed using the recorded NF ground motions and generated ones based on a mathematical method. Finally, numerical experiments and comparative investigations are carried out to comprehensively evaluate the performance of Ep2NgmGAN. The results indicate that the label embedding module is more suitable to deal with the continuous labels and the knowledge-enhanced module makes the model better learn the prior knowledge. In comparison to the representative mathematical methods, the Ep2NgmGAN has much higher efficiency and better or comparable accuracy, making it an appealing tool for NF ground motion synthesis.