Output-only response migration method of single-layer reticulated shells based on generative adversarial network

Abstract

As an important form of important public building roofs, single-layer latticed shells (SLRSs) are equipped with health monitoring systems (SHMs) to collect extensive data to study their complex response characteristics. For SLRSs without structural SHMs, limited response data can only be obtained through on-site measurements, which limits the further application of data-driven machine learning methods on SLRSs. Based on the measured actual response, a traditional generative network can generate fake responses similar to the actual responses to expand the existing response dataset. However, this method is not subject to enough constraints, leading to a small portion of the generated fake responses do not match the actual situation. This paper proposes a response migration method to expand the response dataset of SLRS. Firstly, the output-only response migration method for SLRS is proposed, utilizing frequency domain responses from two different response domains to expand the responses of the actual SLRS through response migration. Then, more constraints are applied to the response migration process to reduce the occurrence of abnormal fake responses through cycle consistency generation adversarial network (GAN). The adaptive response migration generation adversarial network (ARMGAN) is proposed based on cycle consistency GAN, which can automatically select networks with better migration performance. Finally, two numerical spherical SLRS examples and an engineering example are used to verify the proposed method. The responses migrated through ARMGAN have characteristics similar to the actual responses, and ARMGAN can significantly reduce abnormal fake responses.