Synthesizing Dynamic Time-Series Data for Structures Under Shock Using Generative Adversarial Networks
Synthesizing Dynamic Time-Series Data for Structures Under Shock Using Generative Adversarial Networks
复制标题
使用生成对抗网络合成冲击下结构的动态时间序列数据
DOI：
10.1007/978-3-031-04122-8_16
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
Wei, Jie
中科院分区：
文献类型：
--
作者：
Thompson, Zhymir;Downey, Austin R.;Bakos, Jason D.;Wei, Jie
验证状态观测器的高速率结构健康监测需要测试的状态观测器上的一个大型库的预先记录的信号，单变量和多变量。然而，高价值结构的实验测试可能成本和时间过高。虽然有限元建模可以生成额外的数据集，但它缺乏再现信号中存在的非平稳性的保真度，特别是在数字化信号频带的高端。在这项初步工作中，生成对抗网络的研究，为冲击下的电子封装的单变量和多变量加速度信号的合成。生成对抗网络是一类深度学习方法，它学习生成与原始数据在统计上相似但不相同的新数据，从而增强数据的多样性和平衡性。本章介绍了一种方法，用于合成统计上无法区分的时间序列数据的结构下冲击。结果表明，生成对抗网络能够产生与实验测试相似的材料。生成的数据进行统计比较实验数据，并讨论了该方法的准确性，多样性和局限性。
Validation of state observers for high-rate structural health monitoring requires the testing of state observers on a large library of pre-recorded signals, both uni- and multi-variate. However, experimental testing of high-value structures can be cost and time prohibitive. While finite element modeling can generate additional datasets, it lacks the fidelity to reproduce the non-stationarities present in the signal, particularly at the higher end of the digitized signal’s frequency band. In this preliminary work, generative adversarial networks are investigated for the synthesis of uni- and multi-variate acceleration signals for an electronics package under shock. Generative adversarial networks are a class of deep learning approach that learns to generate new data that is statistically similar to the original data but not identical and thus augmenting the data diversity and balance. This chapter presents a methodology for synthesizing statistically indistinguishable time-series data for a structure under shock. Results show that generative adversarial networks are capable of producing material reminiscent of that obtained through experimental testing. The generated data is compared statistically to experimental data, and the accuracy, diversity, and limitations of the method are discussed.